0% found this document useful (0 votes)

118 views126 pages

2023irds Mds

Uploaded by

edmorezy95

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

118 views126 pages

2023irds Mds

Uploaded by

edmorezy95

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 126

INTERNATIONAL

ROADMAP
FOR
DEVICES AND SYSTEMS

2023 UPDATE

MASS DATA STORAGE

THE IRDS IS DEVISED AND INTENDED FOR TECHNOLOGY ASSESSMENT ONLY AND IS WITHOUT REGARD TO ANY
COMMERCIAL CONSIDERATIONS PERTAINING TO INDIVIDUAL PRODUCTS OR EQUIPMENT.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

© 2023 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any
current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating
new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in
other works.
Other trademarks and registration marks are owned by their respective companies.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Table of Contents
Contributors ......................................................................................................................viii
Summary ............................................................................................................................. 1
Solid State .....................................................................................................................................1
HDD...............................................................................................................................................1
Tape ..............................................................................................................................................1
Optical...........................................................................................................................................2
DNA Storage ..................................................................................................................................2
Solid State Storage ............................................................................................................... 3
NAND Flash Storage .......................................................................................................................3
Situation Analysis ................................................................................................................................. 3
Business/Technical Issues .................................................................................................................... 7
Roadmap of Quantified Key Attribute Needs .................................................................................... 13
Critical Issues ...................................................................................................................................... 14
Technology Needs: Research, Development, Implementation ......................................................... 15
Gaps and Showstoppers ..................................................................................................................... 15
Recommendations on Priorities and Alternative Technologies ......................................................... 16
Other Emerging Solid State Memory Technologies ........................................................................ 16
Situation Analysis ............................................................................................................................... 16
Ferroelectric Random Access Memory (FRAM or FeRAM) ................................................................ 17
Magnetoresistive RAM (MRAM) ........................................................................................................ 18
Resistive RAM (RRAM or ReRAM) ...................................................................................................... 19
Comparing the Technologies.............................................................................................................. 21
Storage Class Memory (SCM) ....................................................................................................... 23
Situation Analysis ............................................................................................................................... 23
The Intel/Micron 3D XPoint Memory ................................................................................................. 24
BUSINESS/TECHNICAL ISSUES ............................................................................................................ 25
Gaps and Showstoppers ..................................................................................................................... 26
Recommendations on Priorities And Alternative Technologies ........................................................ 27
Compute Express Link (CXL) ......................................................................................................... 28
Hard Disk Drive (HDD) Technology ..................................................................................... 30
Situation Analysis ........................................................................................................................ 30
Business/Technical Issues ............................................................................................................ 31
Manufacturing Equipment ................................................................................................................. 31
Manufacturing Processes ................................................................................................................... 32
Materials ............................................................................................................................................ 33
Roadmap of Key Attribute Needs ................................................................................................. 37
Critical Issues ............................................................................................................................... 39
Business Vitality.................................................................................................................................. 39
Areal Density Growth Rate ................................................................................................................. 39
Competition from Solid State Drives (SSDs) ....................................................................................... 40
Technology Needs: Research, Development, Implementation ......................................................... 40
Gaps and Showstoppers ..................................................................................................................... 41
Recommendations on Priorities and Alternative Technologies ...................................................... 43
Shingled Magnetic Recording (SMR) .................................................................................................. 43
Sealed Helium Drive ........................................................................................................................... 44
Heat Assisted Magnetic Recording (HAMR) ....................................................................................... 45
Energy assisted Perpendicular Magnetic Recording (EPMR) ............................................................. 47

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Microwave Assisted Magnetic Recording (MAMR)............................................................................ 47

Two dimensional Magnetic recording (TDMR) .................................................................................. 49
Heated dot media............................................................................................................................... 50
Full Disk Encrypting Drives ................................................................................................................. 53
Magnetic Tape Storage ...................................................................................................... 55
Situation Analysis ........................................................................................................................ 55
Business/Technical Issues ............................................................................................................ 57
Roadmap of Quantified Key Attribute Needs ................................................................................ 62
Critical Issues ............................................................................................................................... 65
Technology Needs: Research, Development, Implementation ......................................................... 65
Gaps and Showstoppers ..................................................................................................................... 67
Recommendations on Priorities and Alternative Technologies ......................................................... 68
References................................................................................................................................... 68
Optical Archival Data Storage ............................................................................................. 69
Situation Analysis ........................................................................................................................ 69
Consumer Market .............................................................................................................................. 69
Enterprise Use Case Value Proposition .............................................................................................. 72
Technical Requirements (Business/Technical Issues) ..................................................................... 74
Incumbent Media: Multilayer disc ..................................................................................................... 76
Data Preservation ............................................................................................................................... 80
Optical Storage Systems ..................................................................................................................... 85
Technical Issues ........................................................................................................................... 89
Capacity .............................................................................................................................................. 89
Cost..................................................................................................................................................... 89
Speed .................................................................................................................................................. 89
Potential Solutions (Roadmap of Quantified Key Attributes Needs) ............................................... 89
Future Multilayer Technology ............................................................................................................ 91
Optical Mass Data Storage Technology Roadmap.......................................................................... 93
Emerging Concepts............................................................................................................................. 93
Holographic ........................................................................................................................................ 99
Challenges (Critical Issues) ......................................................................................................... 100
Capacity ............................................................................................................................................ 100
Cost................................................................................................................................................... 101
Speed ................................................................................................................................................ 101
Summary & Key Points ............................................................................................................... 101
Conclusions & Recommendations ............................................................................................... 102
References................................................................................................................................. 102
DNA Data Storage ............................................................................................................ 107
Situation Analysis ...................................................................................................................... 107
Challenges and Issues ................................................................................................................ 109
TCO ................................................................................................................................................... 109
Writing DNA (Synthesis) ................................................................................................................... 112
Reading DNA (Sequencing)............................................................................................................... 115
Sustainability .................................................................................................................................... 117
Biosecurity ........................................................................................................................................ 117
Summary ................................................................................................................................... 117
References................................................................................................................................. 118

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Table of Figures
Figure 1. NAND Flash Shipments ........................................................................................................................................... 3
Figure 2. NAND Flash Shipments ........................................................................................................................................... 4
Figure 3. The Memory/Storage Hierarchy in Cost vs. Performance ...................................................................................... 5
Figure 4. NAND Flash ASP/GB ................................................................................................................................................ 6
Figure 5. SSD ASP per Gigabyte History ................................................................................................................................. 6
Figure 6. Cross-Section of planar floating-gate flash memory cell ........................................................................................ 8
Figure 7. Toshiba's BiCS Vertical NAND Structure (Courtesy of Toshiba Corp (Kioxia)) ........................................................ 9
Figure 8. Radar Chart of Memory Technical Attributes ....................................................................................................... 23
Figure 9. Physical Mock-Up of a 2-Layer 3D XPoint Memory Structure .............................................................................. 26
Figure 10. Stacked Crosspoint Memory Array ..................................................................................................................... 26
Figure 11. Shipped Disk Drive Volumes vs. Time, by Application ........................................................................................ 30
Figure 12. Raw Storage Average Retail Price vs. Time ......................................................................................................... 31
Figure 13. Hard Disk Drive Head Manufacturing Process .................................................................................................... 32
Figure 14. Typical Hard Disk Drive with Key Components Identified (Photo of a Samsung Hard Disk Drive) ..................... 34
Figure 15. Close-up of Actuator Positioning a Head Suspension (HGA) on a Disk (Photo of a WD Drive) .......................... 34
Figure 16. Relationship of Head and Disk Including "Flying Height" of the "Slider" ............................................................ 34
Figure 17. (a) Schematic of TMR head layers, including the Permanent Magnet (PM), Synthetic Anti-Ferromagnetic (SAF),
and Anti-Ferromagnet (AFM) Structures and (b) Air Bearing Surface (ABS) view of a TMR head ..................... 36
Figure 18. Example of a Piezo-Based Head Microactuator .................................................................................................. 36
Figure 19. Schematic Illustration of Magnetic Recording .................................................................................................. 37
Figure 20. 2022 Areal Density Roadmap ............................................................................................................................. 39
Figure 21. Disk Drive Access Density Is Decreasing ............................................................................................................. 42
Figure 22. Shingled Recording and Head Magnetics .......................................................................................................... 44
Figure 23. Western Digital 22TB and Seagate - 20TB Sealed Helium Drives ....................................................................... 45
Figure 24. Curie Point Writing Using Head Assisted Magnetic Recording ........................................................................... 46
Figure 25. Sketch of basic proposed HAMR recording showing head design w/light impinging on the grating ................. 46
Figure 26. HAMR head with the laser source built into the head ....................................................................................... 47
Figure 27. MAMR concept ................................................................................................................................................... 48
Figure 28. Toshiba MAMR/HAMR Roadmap derived by Chris Mellor in The Register Jan. 13, 2022 .................................. 49
Figure 29. Media surface is a two dimensional environment ............................................................................................. 49
Figure 30. Dual reader TDMR concept................................................................................................................................. 50
Figure 31. Comparison of conventional magnetic media to patterned media .................................................................... 51
Figure 32. SEM images of self-assembled patterned magnetic recording media showing servo and address patterns .... 51
Figure 33. Block co-polymer self-assembly ......................................................................................................................... 52
Figure 34. Possible patterned media production process ................................................................................................... 53
Figure 35. (a) Photograph of a 3 module, 32 channel tape head, (b) Illustration of a 3 module 32 channel tape head. The
center reader module (CR) contains 32 data readers (R01-R32) and 2 servo readers (S1 and S2), and the left
(LW) and right writer (RW) modules each contain 32 writers (W01-W32) and two servo readers (S1 and S2).
The read transducers in the reader module are aligned with the write transducers in the writer modules to
enable read while write verification. ................................................................................................................. 56
Figure 36. Illustration of the four data band / five servo band tape layout used in the LTO format. LTO generations 7 to 9
use a 32-channel format illustrated on the right side of the figure in which each data band is subdivided into
32 sub-data bands. 32 tracks are written in parallel with forward wraps (tracks) written in the upper part of
each sub-data band and reverse wraps written in the lower part. Multiple passes back and forth along the
length of tape are required to fill each data band. ............................................................................................ 58
Figure 37. LTO-9 FH Tape Drive and Cartridge .................................................................................................................... 60
Figure 38. (a) IBM Tape Libraries and (b) Spectra Logic Tape Libraries ............................................................................... 61
Figure 39. LTO Consortium Roadmap [8]............................................................................................................................. 64
Figure 40. INSIC Tape Areal Density Roadmap and Areal Density of Tape and HDD Demonstrations and Products [9]. ... 65
Figure 41. Decrease in laser wavelength and spot size with increasing numerical aperture (NA) lenses. . ........................ 70
Figure 42. Music delivery. DVD sales exhibit a similar trend to CD sales with a peak in the year 2004 . ............................ 71

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 43. Storage media comparison. ............................................................................................................................... 72

Figure 44. Accelerated lifetime test results for Sony/Panasonic Archival Disc. .................................................................. 73
Figure 45. BDXL layer structure. ......................................................................................................................................... 77
Figure 46. Land-and-groove recording. Blu-ray Disc on the left vs. AD disc on the right. .................................................. 77
Figure 47. 3rd generation AD multiple-level recording/playback performance .................................................................. 78
Figure 48. Archival disc capacity roadmap. ........................................................................................................................ 79
Figure 49. piqlFilm package ................................................................................................................................................. 80
Figure 50. Digital and analog data on piqlFilm .................................................................................................................... 81
Figure 51. piqlFilm. Reproduced with permission: Piql AS .................................................................................................. 82
Figure 52. Sample binary data on piqlFilm .......................................................................................................................... 82
Figure 53. A piqlReader ....................................................................................................................................................... 84
Figure 54. NETZON HDL 10368 holds 10368 discs enclosed in 864 cartridges and 36 parallel drives. ............................... 85
Figure 55. Sony PetaSite library system. ............................................................................................................................. 87
Figure 56. Panasonic’s Freeze-ray LB-DH6 (left) and LB-DH7 (right) data archiver systems. ............................................. 88
Figure 57. Folio Photonics manufacturing process and product. ....................................................................................... 91
Figure 58. Guide layer tracking scheme. ............................................................................................................................. 92
Figure 59. Scheme of Folio Photonic’s confocal fiber focus error signal scheme for fluorescent signals. .......................... 92
Figure 60. 5D storage of “The Hitchhiker’s Guide to the Galaxy.” a) Illustration of data encoding and decoding. b) The
birefringent images of data voxels of different layers. Inset is the transmission of 100-layer data in the visible
range. c) The birefringent images after removing the background of (b). Insets are enlargements of small
region (10 μm x 10 μm). d) Polar diagram of the measured retardance and azimuth of all voxels in (c). ........ 95
Figure 61. Sample DOTS data encoding mechanism. .......................................................................................................... 97
Figure 62. Cerabyte writing scheme .................................................................................................................................... 98
Figure 63. Cerabyte’s storage concept ................................................................................................................................ 98
Figure 64. Cerabyte’s product roadmap of enterprise archiving systems ........................................................................... 99
Figure 65. The Zone of Potential Insufficiency .................................................................................................................. 107
Figure 66. DNA Double Helix: National Human Genome Research Institute (NHGRI) ...................................................... 108
Figure 67. DNA Density; Preserving our Digital Legacy: An Introduction to DNA Data Storage; DNA Data Storage Alliance
[4] ..................................................................................................................................................................... 108
Figure 68. The Evolving Storage Pyramid .......................................................................................................................... 110
Figure 69. Half-life of verious DNA preservation methods. 5............................................................................................. 111
Figure 70. Estimated Cost of Writing and Storing – Legacy vs. DNA4 ................................................................................ 112
Figure 71. Write latency and throughput for various storage solutions: DNA Data Storage Alliance ............................... 112
Figure 72. 2022 IARPA Roadmap for DNA synthesis, courtesy David M. Markowitz. Assumes single stranded DNA, 150
nucleotides in length (20 nt flanking primers), encoded at 1 bit/nucleotide. ................................................. 113
Figure 73. Electrochemical DNA synthesis on a nanoscale array. (c) An overview of the nanoscale DNA synthesis array
with scanning electron microscopy images of the 650-nm electrode array and enlarged view of one
electrode. (e) Illustration of the wells patterned with ssDNA oligos with multiple copies of each oligo per
synthesis location.11 ......................................................................................................................................... 114
Figure 74. Sequencing by Synthesis: Illumina .................................................................................................................... 115
Figure 75. Nanopore Sequencing: National Human Genome Research Institute (NHGRI) ............................................... 115
Figure 76. Cost per Raw Megabase of DNA Sequencing, DNA Sequencing Costs: Data from the NHGRI Genome
Sequencing Program (GSP) Available at: www.genome.gov/sequencingcostsdata ........................................ 116

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Table of Tables
Table 1. NAND Flash Chip Roadmap .................................................................................................................................... 14
Table 2. Solid State Memory Technology Comparisoni ....................................................................................................... 22
Table 3. Magnetic Mass Data Storage Technology Roadmap – HDD .................................................................................. 37
Table 4. Magnetic Mass Data Storage Technology Roadmap – Tape .................................................................................. 63
Table 5. Common Enterprise Archival Storage Media ......................................................................................................... 74
Table 6. Summary of select current and novel optical products/technologies ................................................................... 76
Table 7. Main parameters of AD discs. Reproduced with permission: Sony Group Corporation and Panasonic Holdings
Corporation ....................................................................................................................................................... 80
Table 8. Amethystum ZL series BDXL Optical Storage System Specifications ..................................................................... 86
Table 9. Sony PetaSite library system specifications. Reproduced with permission: Sony Group Corporation ................. 87
Table 10. Freeze-ray specifications ..................................................................................................................................... 89
Table 11. Optical Mass Data Storage Technology Roadmap – BDXL, Archival Disc, and Folio Photonic Disc ..................... 93

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Contributors
Tom Coughlin, Chair
Roger F. Hoyt, co-chair

Name Affiliation
Ed Childers IBM
Tom Coughlin Consultant
Ron Dennison Consultant
Simeon Furrer IBM
Roger Hoyt Consultant
John Hoffman Ernst&Young/DNA Data Storage Alliance
Dave Landsman Western Digital/DNA Data Storage Alliance
Mark Lantz IBM
Kevin Lu Folio Photonics
Niranjan Natekar Western Digital
Ken Singer Folio Photonics
Doug Wong Kioxia

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Summary
SOLID STATE
Over the past 30 years (1993 to 2023), the NAND flash market has grown from zero to become a nearly
$60B dollar market not only by NAND displacing existing storage media, but also by NAND-based
products enabling new markets. The initial market for NAND flash was audio tape replacement in
digital telephone answering machines, but the market that jump started NAND flash adoption was its
use in digital cameras. A proliferation of small form flash memory cards followed the advent of digital
photography: PCMCIA (PC Cards), CompactFlash, Smart Media, Multi Media card, and Secure Digital
(SD) cards. The demand for flash grew along with the digital camera market. As the cost of NAND
flash fell, the market grew as floppy disks and writeable CD’s began to be replaced by USB flash
drives. The transition from audio tape and CD players to digital audio MP3 players was also enabled
by the falling cost of NAND flash. At what point alternative memory technologies like MRAM,
FeRAM, ReRAM, or others, might displace DRAM or NAND is unknown. Today, all of these
alternative technologies are still more costly per bit than DRAM or NAND flash, and this prevents
them from being selected as replacements except in those rare circumstances where the cost is less
important than certain important attributes they provide.

HDD
Of current mass data storage technologies, in terms of storage capacity shipped, hard disk drives
(HDD) are by far the largest single component of the mass data storage industry. Today the HDD
market continues its decline in unit volume primarily due to displacement by solid state drives.
However, the demand for data center nearline storage continues to grow and technology advances
such as helium filling, more heads and disks, dual actuators and heat assisted magnetic recording
(HAMR) have fueled continuing capacity increases. Seagate is now shipping 32 TB drives and
anticipates 50 TB drives in 2026. Future capacity growth will depend on the further development of
HAMR as well as new technologies such as next generation TDMR and heated dot magnetic
recording.

TAPE
The continued exponential growth in the creation of digital data combined with the recent slowdown
in areal density and capacity scaling of hard disk drives is driving demand for cost effective data storage
solutions. Magnetic tape technology is particularly well suited to help meet this demand due to its very
low total cost of ownership and its efficient data center footprint that is approaching 5PB/ft 2. Part of
the TCO advantage of tape arises from its very low power consumption, which also contributes to its
small CO2 footprint. The natural physical airgap of tape solutions also provides an additional level of
protection against accidental data deletion and cyber security threats. Moreover, recent tape areal
density demonstrations provide confidence that tape has the potential to continue scaling areal density
for multiple future generations with cartridge capacities expected to reach hundreds of TB per cartridge
within the next decade. All these benefits have resulted in an increased adoption of tape, particularly
among hyperscale cloud companies, and are driving growth of the tape market.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

OPTICAL
Optical storage media is undergoing a shift from its traditional role in consumer media distribution to
a focus on enterprise and institutional archival storage. To enhance capacity while minimizing costs,
emerging optical technologies are exploring storage solutions in the third dimension and beyond.
Robust library systems tailored for optical media are emerging to meet the stringent requirements of
enterprise-level storage demands.

The low maintenance and operating energy costs of optical media, coupled with infrequent remastering
needs, position it as a naturally advantageous solution in the sustainability-conscious landscape of data
archiving and preservation. Optical Write-Once, Read-Many (WORM) technologies, with their air-gap
feature, offer cybersecurity advantages. In terms of energy consumption, optical technologies exhibit
the lowest levels both intrinsically and in the context of data center environmental control, promising
significant reductions in greenhouse gas emissions.

Despite these benefits, critical challenges persist in the optical data storage domain, including the need
to lower initial costs, expand capacity, increase speed, and improve error management. Exploring
possibilities such as multidimensional media, remote-write libraries, femtosecond lasers, and high-
speed display and imaging technologies opens new opportunities for enterprise optical data storage.

DNA STORAGE
The world is attempting to digitize unprecedented amounts of information. This information can be
valuable if mined, stitched together, or otherwise searched and analyzed; however, the cost of saving
the massive amount of data associated with this information is beginning to overwhelm the ability to
pay for it using conventional storage technologies. This trend is leading system designers to look for
new storage technologies which can sustain the densities, access flexibility, and TCO needed for this
wave of digitization. One of the candidate technologies being considered is synthetic DNA. DNA is
a potentially compelling storage medium due to its ~1bit/nm3 bit size, the fact that it is incredibly stable
at room temperature if kept dry, and that it can be read in the future even if the original writing/reading
technology no longer exists. This combination of factors enables the potential of the proverbial
“datacenter in a shoebox” as compared to incumbent storage technologies: small footprint, low power,
no fixity checks or technology refresh. In other words, compelling TCO.

While DNA data storage is not yet ready for productization today, with the huge advances in
medical/scientific DNA technology and applications over the past several decades, plus academic
research and commercial biotech both targeted at DNA data storage over the past decade, the basic
foundations for synthetic DNA as a data storage medium have been demonstrated[17]. It is thus
reasonable to expect that a path to a synthetic DNA data storage ecosystem will come into focus over
the next 5-10 years.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Solid State Storage

NAND FLASH STORAGE
SITUATION ANALYSIS
Over the past 30 years (1993 to 2023), the NAND flash market has grown from zero to become a
nearly $60B dollar market not only by NAND displacing existing storage media, but also by
NAND-based products enabling new markets. See also the IRDS NVM Technology Roadmap at
this site: https://irds.ieee.org/images/files/pdf/2022/2022IRDS_MM_Tables.xlsx.

The initial market for NAND flash was audio tape replacement in digital telephone answering
machines, but the market that jump started NAND flash adoption was its use in digital cameras.
A proliferation of small form flash memory cards followed the advent of digital photography:
PCMCIA (PC Cards), CompactFlash, SmartMedia, MultiMedia card, and Secure Digital (SD)
cards. The demand for flash grew along with the digital camera market. As the cost of NAND
flash fell, the market grew as floppy disks and writeable CDs began to be replaced by USB flash
drives. The transition from audio tape and CD players to digital audio MP3 players was also
enabled by the falling cost of NAND flash. Figure 1 shows NAND flash shipment revenue from
2013 through 2022 and estimated for 2023.

NAND flash shipments (billions of dollars)

0
2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023

Figure 1. NAND Flash Shipments

Source: Forward Insights

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Other products enabled by the high density and low bit cost of NAND include portable GPS
devices and Personal Digital Assistants (PDAs), but the next big application was the creation of
the smart phone, and that application continues to drive a large segment of the NAND flash market.
It is only within the last decade that NAND flash has become inexpensive enough to start
displacing traditional rotating media (i.e., hard disk drives) and today (2023), solid state drives
(SSD) are the largest market for NAND flash memory. Figure 2 shows NAND flash byte
shipments from 2013 through 2022 and estimated for 2023.

NAND Flash bytes shipped (billions of GB)

1,000.0
900.0
800.0
700.0
600.0
500.0
400.0
300.0
200.0
100.0
0.0
2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023

Figure 2. NAND Flash Shipments

Source: Gartner

As flash-based SSD costs have fallen, HDDs have been displaced: first in consumer PCs, and
increasingly in data centers, as SSDs have come to occupy the tier of frequently, and randomly,
accessed data. SSDs and HDDs continue to coexist because HDDs (and tape) will continue to
offer the lowest cost per bit for the foreseeable future, but the designers of storage systems and
servers recognize the benefit of SSDs for improving data access time and reducing power
consumption, and the development of SSD form factors specifically for this market indicates this.

Figure 3 shows a view of the memory and storage hierarchy comparing cost and performance on
a log-log scale. DRAM and SRAM (L1, L2, L3) are volatile memories while NAND, HDD, and
tape are non-volatile. The key point is that memory performance is correlated with cost per bit.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

107
1.E+07

10 6
1.E+06
L1

10 5
1.E+05 L2
Bandwidth (MB/s)

104 L3
1.E+04

103
1.E+03
DRAM

NAND

10 2
1.E+02
HDD

101
1.E+01
Tape

10 0
1.E+00
10 -1
1.E-01 10 0
1.E+00 10 1
1.E+01 10 2
1.E+02 10 3
1.E+03 10 4
1.E+04 10 5
1.E+05 10 6
1.E+06

Price per Gigabyte

Figure 3. The Memory/Storage Hierarchy in Cost vs. Performance

Source: Objective Analysis

While NAND flash is sometimes directly connected to an SoC (system on chip) or microcontroller,
it is most often used with a controller chip specifically designed to support a specific interface. An
SSD controller might support PATA, SATA, SAS or NVMe. A controller might be packaged in
the same package as the NAND flash itself, as is the case for eMMC (embedded MultiMedia Card)
or UFS (Universal Flash Storage).

Due to cost and capacity, SSDs exist as a faster, but more expensive, storage tier. HDDs are used
for bulk mass storage and SSDs are used for speed. Frequently requested and typically randomly
accessed data resides in the SSDs, while less-frequently used and typically sequentially accessed
data is kept in high-capacity HDDs.

When measured in cost per gigabyte (GB), an SSD is more expensive than an HDD. The cost/GB
gap between SSD and HDD has been shrinking over the past decade but is unlikely to cross due
to the fact that although NAND bit density per die continues to increase, so does the areal bit
density of HDDs. Figure 4 shows the average selling price of a GB of NAND flash memory from
2013 through 2022 and estimated for 2023. Figure 5 shows the average selling price of a GB of
memory in an SSD from 2013 through 2022 and estimated for 2023. SSD prices are somewhat
higher than the price of raw NAND flash memory. Because of this, SSDs first replaced those
HDDs that were being used in a way that increases I/O speed at the expense of capacity.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

NAND Flash average Selling Price per GB

0.700

0.600

0.500

0.400

0.300

0.200

0.100

-
2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023

Figure 4. NAND Flash ASP/GB

Source: Gartner

In the past, storage systems used a few HDDs in a faster storage tier that used “short stroking” or
“destroking” to provide faster data access. In a short-stroked HDD, the system used only the tracks
on the outside edge of the platters, limiting actuator travel distance and accessing the physical
media where data passes most rapidly under the heads, thus increasing overall IO performance.
Today’s redesigned storage systems have abandoned short-stroked HDDs, replacing a number of
short-stroked HDDs with a single SSD. This increased performance while lowering costs.

SSD ASP/GB
1.000
0.900
0.800
0.700
0.600
0.500
0.400
0.300
0.200
0.100
-
2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023

Figure 5. SSD ASP per Gigabyte History

Source: Gartner

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Five years ago, SSD penetration percentages in consumer or client laptops was still relatively low.
The reason was higher cost and no appreciable performance improvement using the SATA
protocol. Today, the SSD penetration rate exceeds 95% due to the falling cost of SSDs as well as
the performance improvement gains and reduced boot times of using NVMe (Non-Volatile
Memory Express). NVMe was designed to take advantage of the shorter latency associated with
solid state non-volatile memory (not just NAND flash, but also any future non-volatile memory
technology).

BUSINESS/TECHNICAL ISSUES
There are many parts of the flash memory value chain, some of which are more profitable than
others. Flash card makers and many SSD manufacturers compete by perfecting manufacturing
efficiency and inventories, and through good responsiveness to the market, but this market is
undifferentiated so margins are low. Other SSD makers use proprietary controller chips, or other
differentiators to justify a higher price for their product. In both flash cards and SSDs the flash
memory tends to account for 80% or more of the bill of materials, so changing the controller or
other non-flash portions of the system may result in a high payback. Makers of higher-value
devices including smart phones, and tablet PCs can reap larger profits (from storage) thanks to the
greater differentiation of these products.

In undifferentiated markets the bulk of the profits and the greatest challenge lay in the production
of the NAND flash chips used to make those products rather than in other technologies and
manufacturing capabilities.

Intellectual property is an important part of the flash memory business. Flash manufacturers invest
heavily in their intellectual property portfolios and patent protection has been very useful in
assuring that this research effort is properly rewarded. Prior to its acquisition by Western Digital,
SanDisk signed royalty agreements with nearly all flash chip, card, and controller manufacturers
to generate a royalty stream that offset a significant portion of the company’s R&D cost.

Over its lifetime NAND’s price per gigabyte has declined rapidly, allowing the technology to
displace entrenched storage solutions including photographic film, floppy disks, magnetic tape,
and rotating optical media. Today SSDs and video applications consume the largest per-device
quantity of NAND flash storage capacity.

NAND flash chips are manufactured by Samsung and SK Hynix in Korea, WD and Kioxia in a
joint venture in Japan, Micron in the US and Singapore, YMTC in China, and Macronix and
Winbond in Taiwan.

MANUFACTURING EQUIPMENT
NAND flash chips are manufactured using standard semiconductor processing equipment. Cost
is a key focus, so NAND manufacturers rapidly migrate process technologies and moved from
200mm wafers to 300mm wafers starting in 2007. In fact, the process migration of NAND flash
has become so important that flash became the process driver for those companies who
manufacture both flash and another technology (i.e. DRAM). A “process driver” is the technology
used to perfect the manufacturing process of a company’s most advanced and lowest-cost
technology. However, the move from planar 2D NAND to vertical 3D NAND has changed the

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

need to create the smallest sized features and instead, the availability of advanced deposition and
etch technologies are driving the development of the latest 3D NAND designs.

Flash is manufactured using a production process similar to that of standard CMOS logic, a process
used to manufacture most semiconductors. Historically, the main difference between flash and
CMOS logic was that flash added a thin gate oxide layer and a polysilicon floating gate to the
process to enable the construction of a floating gate. This floating gate, the key to any electrically
programmable nonvolatile memory, can be programmed either by hot electron injection or by
tunneling, both of which force electrons through the thin gate oxide insulator. Even though flash
adds only a small number of additional layers to a standard CMOS process, the development of a
flash process is quite complex, and provides a barrier to entry into this market. Figure 6 shows
cross-sections of the floating-gate memory cell in reading, programming and erasing. This cell is
programmed using tunneling, rather than hot electron injection.

Figure 6. Cross-Section of planar floating-gate flash memory cell

Left: Reading, Middle: Programming, Right: Erasing
(Source: Samsung Semiconductor Company)

While floating gate NAND flash memory continues to be manufactured, more than 90% of the
NAND flash memory bits produced today are 3D NAND flash memory. The industry has
transitioned from the 2D planar process, in which the memory array lies on the surface of the
silicon wafer, to a vertical 3D process in which memory cells are formed in stacked layers. No
more development is being done using the 2D planar process.

The reason for the transition is to increase the number of bits per die, which is necessary in order
to reduce bit cost. Planar 2D flash has undergone continuous cost decreases by following Moore’s
Law by shrinking the size of the transistors every 18 months or so. But this constant shrinking
eventually brought flash to the limits of scaling. As the memory cell transistor got smaller, the
number of electrons it stored also got smaller, which reduced write/erase cycle endurance and data
retention time. When the lithography node hit the mid-teens (15nm), it was no longer feasible to
shrink further. At this point, the die density was approximately 128 Gbits for an MLC (2 bit per
cell) device.

At this point, the NAND flash market needed to transition to a new technology. 3D NAND flash.
3D (three-dimensional) NAND (Figure 7) stacks bit cells rather than shrinking them. The number
of bits per chip increases in proportion to the number of layers. Although this process does scale
up in total chip density each generation, it is less cost effective than the density increase achieved

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

by a lithography shrink in the past. This 2007 Toshiba invention describes a vertical NAND
structure using conventional semiconductor materials but a relatively complex process using
several steps that have never previously been used in chipmaking.

Figure 7. Toshiba's BiCS Vertical NAND Structure (Courtesy of Toshiba Corp (Kioxia))
Each manufacturer now uses a different variant of this process. Samsung has been shipping 3D
NAND-based SSDs since July 2014. Other companies delayed their launch of the technology until
they saw a clear path to profitability, but as of 2023, NAND vendors have now been in mass
production of 3D NAND for several generations and 3D NAND comprises >90% of the total
NAND flash market in bits.

The conversion from yesterday’s planar to today’s 3D process was a significant transition since
the manufacturing equipment mix is different between the two. Planar NAND uses a mix that is
higher in the need for lithography tools, while 3D NAND production uses more deposition and
etch tools. Historically, flash manufacturers have been able to migrate from one process to the
next through the addition of incremental numbers of new tools that are absolutely necessary to the
process migration, using already-installed tools for all of the other steps. The equipment mix
needed for 3D NAND was sufficiently different that migration could only be performed through
the addition of significantly more tooling than in previous process migrations.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

In the last iNEMI report, it was speculated that 3D NAND would eventually take over the market
from the existing 15nm planar NAND flash. This has now occurred and 3D NAND bits represents
most of the NAND market.

Now that 3D NAND is the mainstream process, the challenge becomes how to continue the path
toward more bits per die and lower bit cost. The accepted way to continue cost reductions with
3D NAND is to add layers to the chip. Manufacturers started the first generation of 3D NAND at
24-36 layers. For a number of years flash makers thought that the number of layers would be
limited to about 100 layers due to the difficulties of etching a deep enough and narrow enough
hole to build the columns in Figure 7. In 2015 SK Hynix revealed a solution to this problem called
“String Stacking” which builds a chip of perhaps 100 layers, then builds another 100-layer chip on
top of that, and so forth. There is no clear limit on how high such a stack can be practically made,
but it is clear that the 3D NAND architecture has many years of life ahead of it.

There will be a point in the future in which current 3D NAND technology can no longer scale cost
efficiently to higher density chips. While areal density per die increases with increasing layer
count, there is also an incremental additional cost per layer. In the era of 2D floating gate NAND,
lithography shrinks directly enabled smaller memory arrays and resulted in lower bit cost.
However, bit cost reduction is more challenging in the 3D era because feature sizes are not
shrinking at the same rate. At some point, there is the expectation that alternative technologies
will eventually replace flash, but it is likely for any future memory to still be composed of a 3D
physical interconnection of transistors and have layer count limitations for the manufacturing
process.

MANUFACTURING PROCESSES
The NAND flash market is very cost sensitive and capital intensive. NAND manufacturers can
effectively compete only by using a manufacturing process on a par with that of their competition.
Process advancements can only be achieved through the use of the most expensive tooling, and
this limits the market to a handful of key players who have access to billions of dollars of capital.
As a rule of thumb, a manufacturer will need to upgrade its wafer fabrication plants from one
process node to the next every 2 years.

Process line widths are measured in nanometers (nm), or 10-9 meters. NAND manufacturers have
migrated the vast majority of their production from a planar mid-teens nm process to the 3D NAND
process. While first generation 3D NAND products started at 24-36 layers, today, the highest layer
count 3D NAND currently shipping is 232 layers.

MATERIALS
The materials used to manufacture NAND flash memories are common to nearly all other
semiconductor processes. Semiconductor manufacturing requires extraordinarily pure materials,
since the key to making a semiconductor is to manipulate small impurities within extremely pure
silicon crystals. Typical materials include raw silicon wafers, photographic emulsion and
developers, high-purity gases, including hydrogen, oxygen, nitrogen, and silane, aluminum and
copper sputtering targets, boron, arsenic, tantalum, and other dopants, de-ionized water, and an
abundant dependable supply of electricity. Some processes use wet steps that depend upon a supply
of acids and other reactive liquids.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

From a materials standpoint the 3D NAND process is driving relatively minor changes since most
manufacturers need to convert from a conductive polysilicon floating gate to a charge trap, most
commonly made using silicon nitride (which is already used as a passivation layer over the top of
all silicon semiconductors), and in some cases this will be joined by the replacement of a
polysilicon top gate with tantalum and the gate dielectric with alumina. Silicon nitride, alumina,
and tantalum are regularly used in semiconductor processing and are well understood and
abundantly available.

Memory makers anticipate that one or more of the new “emerging” nonvolatile technologies that
are currently being researched will become attractive at some point in the future. These
technologies include magnetic materials, ferroelectrics, chalcogenide glasses, and other materials
that will be added to the silicon process. Some of these technologies are already in production.
Although these new technologies may not constitute an important part of today’s market, they
could come into widespread use within the timeframe of this roadmap’s outlook.

One issue with these new technologies is that they introduce a new material into the process, and
new materials always create new problems in the fabrication plant since they are not as well
understood as those technologies that have been in volume production for a long time. For this
reason there will be false starts as these newer technologies vie to replace established silicon-only
semiconductor processes.

QUALITY/RELIABILITY
NAND flash as a raw storage medium has many idiosyncratic behaviors and failure modes, so a
controller is always used with NAND-based devices to perform error correction and manage wear-
out. The quality and reliability of storage devices that use NAND flash (i.e., SSDs, flash cards,
USB flash drives) is measured through three main indices:

• Data integrity as provided by the controller,

• Memory cell wear or endurance
• Lifetime of the data in the device (data retention)

Each of these will be addressed in order:

Data integrity: The cost per gigabyte of NAND flash is lower than that of any other semiconductor
memory. NAND’s low cost is achieved by trading off price against data integrity: the lower the
data integrity, the lower the price, so NAND has been designed in a way so that bit errors are
anticipated and allowed to occur. Advanced forward error correction is required to use NAND
flash. Bit errors are corrected external to the flash chip in a controller chip that uses the same sort
of error correction code (ECC) approaches used to recover bit errors in hard disk drives. This is a
very well understood technology. In today’s 3D NAND, LDPC is now the most commonly used
ECC with a correction strength of approximately 120-160 bits/1kByte.

All NAND chips include an area for the storage of parity bits (ECC) that coexist with the data
array in order to enable error correction. Controller designers differentiate their products by using
these parity and coding bits in different ways. Although some companies pride themselves on
having algorithms that provide higher data integrity than their competition, all modern algorithms
provide data that users accept as error-free. Users are comfortable that SSDs, flash cards, and USB
THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023
COPYRIGHT © 2023 IEEE. ALL RIGHTS RESERVED.
12

flash drives will accurately store either code or data, and no single supplier has had to overcome a
reputation of supplying media that does not accurately replicate a file.

Another concern with NAND flash is the wear mechanism. All EEPROM (electrically erasable
programmable read only memory) technologies (i.e. NOR flash and NAND flash) exhibit wearout
mechanisms that put a limit on the number of write/erase cycles sustainable. This characteristic,
called “endurance”, varies with the number of bits stored in a memory cell. NAND flash chips are
categorized in terms of how many bits can be stored in a cell:

• Single Level Cell (SLC), which stores one bit per transistor
• Multi-Level Cell (MLC), which stores 2 bits per transistor
• Triple Level Cell (TLC), which stores 3 bits per transistor
• Quadruple Level Cell (QLC), which stores 4 bits per transistor
• Penta Level Cell (PLC), which stores 5 bits per transistor

TLC (as well as QLC and PLC) terms a little inaccurate because there are actually more levels
than implied, but the names have stuck. In SLC, there is a single threshold voltage that
differentiates between a 0 and a 1 state. In MLC, there are 3 threshold voltages to distinguish
between the 4 states necessary to store 2 bits of information. But for TLC, there are more than 3
“levels” needed to differentiate the 8 charge states necessary to store 3 bits.

The more bits per cell that a NAND chip can store, the fewer write/erase cycles it will be able to
sustain. A typical SLC NAND flash chip made in a 24 nm planar floating gate process may be able
to withstand 50k-100k write/erase cycles while a 15nm MLC chip is rated to only 3k cycles. The
transition to 3D has allowed the individual cells to get larger again, so a TLC 3D NAND cell can
also achieve 3k cycle endurance. QLC and PLC cells of the same size achieve fewer cycles.

The controller that manages the flash in an SSD, card, or USB flash drives detects endurance
failures and either corrects the failed bits or maps the blocks that contain excess failed bits out of
the device, disallowing their subsequent use. This is possible as long as there are spare blocks
available. Flash controllers attempt to spread the wear out across all blocks (wear leveling), so
that all blocks have a similar amount of wear. Sophisticated SSD controllers also perform write
coalescing and write gathering to reduce the number of times a block is written to. Regardless of
how well all of these techniques work, however, at some point, the flash in a device will be worn
out.

Data retention: The charge in a NAND flash determines the bit values. This charge can be expected
to leak over time and temperature, creating bit errors that eventually become numerous enough
that uncorrectable read errors occur.

NAND flash chips are typically specified to retain data for 10 years when new and typically 1 year
at the end of life (i.e. the maximum rated write/erase cycles), when powered down. This compares
against 5 years for magnetic media, and up to 100 years for recordable optical media. JEDEC
(Joint Electron Device Engineering Council) published a number of specifications enabling the
standard measurement and characterization of endurance and data retention in NAND-based chips
and devices. As a chip experiences more cumulative write/erase cycles, the data retention time

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

will decrease. Such failures are the result of lattice disruptions in the insulating oxide which create
leakage paths. As the device experiences more cumulative write/erase cycles, it develops more
lattice disruptions, increasing this leakage. This is the end of life wearout mechanism for all types
of flash and results in insufficient data retention time.

Unlike magnetic media, NAND is unaffected by extraneous magnetic fields. NAND flash is less
sensitive to heat than optical and magnetic media, and is often specified to operate at up to 85ºC
and to withstand storage temperatures of up to 150ºC.

ENVIRONMENTAL TECHNOLOGY
Every semiconductor wafer fabrication plant generates hazardous waste, and the handling of that
waste is counted among the many criteria that impact the design and operation of a plant. Gas
emissions are typically cleaned by scrubbers on a facility’s roof. Wastewater is often purified and
recycled within the facility, to avoid concerns of contaminating groundwater, streams, or other
bodies of water.

All semiconductor manufacturers perform on-site decontamination of factory effluents and have
regular programs of hazardous waste removal. The contaminants removed from effluents are
sealed in drums and conveyed by bonded and insured carriers to hazardous waste dumps.

Semiconductor production generates a low volume of such hazardous wastes, and despite the high
costs of removing it, the cost factor of such handling does not place a material impact on overall
costs. Should a fab in one country not be held to the same environmental standards as those that
apply in a competing country, it would not give that fab a meaningful cost advantage over its
competition.

TEST, INSPECTION AND MEASUREMENT (TIM)

One important challenge for flash manufacturers is the implementation of sufficient test
procedures. Historically, NAND chip densities (capacities) have doubled each generation. and the
time required for testing increase in proportion to the density of the chip. To prevent this from
impacting manufacturing costs, NAND flash manufacturers have started to apply statistical
methods to avoid having to test every bit in the device.

Internal self-test mechanisms are also used, and will increase in sophistication over time. These
internal test mechanisms will allow less test equipment to be used to test more chips. Although
this approach will not help reduce the WIP costs of testing, it will cut the capital expenditures
required in this area.

ROADMAP OF QUANTIFIED KEY ATTRIBUTE NEEDS

The data in Table 1 is based on the IEEE IRDS Semiconductor Roadmaps (International Roadmap
for Devices and Systems) to estimate where NAND flash is headed over the next several years.
This roadmap can be found at https://irds.ieee.org/

Table 1 shows a prediction of how NAND flash processes develop over time and the transition as
the industry has moved from a 2D planar process to a 3D NAND structure.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Table 1. NAND Flash Chip Roadmap

2013 2015 2017 2019 2021 2023 2025 2027 2029
Density 64Gb 128Gb 256Gb 256Gb 512Gb 1Tb 1Tb/2Tb 2Tb/4Tb 4Tb/8Tb
(MLC) (MLC) (TLC) (TLC) (TLC) (TLC) (TLC+) (TLC+) (TLC+)
Planar 19nm 15nm N/A N/A N/A N/A N/A N/A N/A
Process
3D Layers 32-48 64-96 112- Low High 300+ 500+
176 200s 200s

While the development of new planar 2D NAND processes has ended, the 3D era of NAND flash
is still maturing. Just a few years ago, it was thought that 100 layers might be a practical limit, but
in 2023, 3D NAND with layer counts exceeding 200 layers is now in production. TLC 3D NAND
is mainstream today, but the percentage of QLC continues to grow.

CRITICAL ISSUES
LIMITATIONS OF THE FLASH PROCESS
Flash makers have for many years understood that standard planar flash designs would not be able
to shrink below some certain process geometry. NAND flash manufacturers generally expected
that the last process generation at which standard NAND could be manufactured was around 15nm
and this prediction turned out to be accurate. The main problems were cross-coupling between
adjacent cells, and decreasing data retention time due to the decreasing number of electrons that
could be stored in each memory cell transistor.

For the 3D stacked cell, the physical memory cell transistor is significantly larger than in the 2D
process. The memory transistor is now a cylindrically shaped cell that can store more charge than
the 2D cell and made possible the transition from 2 bits per cell to 3 bits per cell. QLC is also now
available and demonstrations of 5 bit per cell (PLC) have also been done. It now seems possible
to stack hundreds of layers in tiers to achieve high areal density per die. For each additional layer
in the 3D NAND, there will be an incremental increase in die density, but also an incremental
increase in wafer cycle time and in cost per wafer.

The path to continuing to increase the die density and reduce the cost per bit will involve:
minimizing the growth in layer count, increasing the density of memory cells per layer (increasing
areal density per layer), decreasing the size of the memory holes, increasing the number of bits
stored per transistor (TLC to QLC to PLC), and maintaining a uniform high aspect ratio etch for
each memory hole. These are design and manufacturing issues and must be solved to maintain the
good yields necessary for profitable manufacturing.

Ultimately, there will be a limit to the number of layers that 3D NAND can practically achieve,
but it will be some time before that limit is reached. As of 2023, projections of up to 1000 layers
has been discussed. Any alternative memory technology that could potentially replace 3D NAND
in the future will be faced with similar manufacturing issues. On a 3D NAND with 128 layers, the
depth of the memory hole (120nm in diameter) was reported to be 6-8 microns or about 55nm per
layer. On a recent 200-layer 3D NAND, the stack was reported to be 5.5 microns or about 20-
30nm per layer.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

A practical problem with the increasing density of 3D NAND flash die each generation is the
reduction in the number of dies necessary to create an SSD of a given capacity and the performance
limit per die. Using a 1Tb TLC die (128GB per die), only 2 die are needed to create a 256GB
SSD. The performance of this SSD will be bottlenecked by the lack of parallel NAND channels
to the SSD controller as well as the speed of each NAND channel. Work is now taking place in
JEDEC to improve the interface speed and performance of future NAND flash devices.

COMPETING TECHNOLOGIES IN THE SHORT TERM

The single most important enabler for success in the major semiconductor memory markets is low
cost per bit. In the semiconductor memory market for discrete external memories, DRAM and
NAND flash are dominant and are expected to remain so for the next decade due to economies of
scale.

NAND flash memory continues to make solid inroads into the established mass storage markets.
SSDs and all flash arrays are popular as a fast storage layer between HDDs and DRAMs, and SSDs
have replaced high spindle-speed HDDs (10k & 15k RPM). However, the ultra-high capacity
HDD market will remain unavailable to flash for the foreseeable future due to its lower cost per
bit.

TECHNOLOGY NEEDS: RESEARCH, DEVELOPMENT, IMPLEMENTATION

The IRDS is carefully evaluating the technologies that will become necessary to allow NAND
flash to continue to be viable through the next decade. This technology development will be
collectively funded by all participants in the semiconductor memory business.

Silicon is very well understood and this gives the material a great cost advantage over competing
technologies. Silicon always gains the upper hand in manufacturing costs.

GAPS AND SHOWSTOPPERS

Over the past couple of decades, researchers expressed a belief on several occasions that
insurmountable obstacles to scaling flash memory were at hand. Once that obstacle was reached,
other researchers have succeeded in devising some ingenious method of sidestepping it. This has
happened for at least four planar NAND process generations, and when planar NAND flash finally
did hit the wall, it was replaced by 3D NAND flash which will enable NAND flash technology to
continue to dominate the non-volatile memory market for at least the next decade.

The cost of a gigabyte of NAND flash memory has continually decreased at an average rate of
30% annually since its invention. In 2002 NAND flash transitioned from SLC to MLC, doubling
the bits per cell. Volume production of 2D TLC commenced in 2010 with TLC being the dominant
type of 3D NAND today. Some four-bit QLC NAND has shipped in the past, but this technology
is expected to become a significant portion of the market in the future thanks to the robustness of
today’s 3D NAND flash technology.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

RECOMMENDATIONS ON PRIORITIES AND ALTERNATIVE TECHNOLOGIES

A very high level of ongoing capital spending is necessary in the flash business. NAND suppliers
must be willing to continue to invest even when the market suffers a price collapse. This is the
recipe for success in any undifferentiated semiconductor market.

High-capacity data storage is experiencing incredible demand growth due to cloud providers and
hyperscale data centers. The key metric has always been total cost of ownership, but performance
and energy consumption per rack. While HDD continues to offer a lower cost per GB, the gap
continues to shrink.

3D NAND flash also the storage requirements of mobile devices include low power drain, small
size and low cost. Compact, inexpensive, efficient, high-density nonvolatile storage is critical for
smart phones, tablets, laptops and other portable consumer devices.

Today’s most promising alternative technologies are MRAM (Magnetic Random Access
Memory), which is viewed as a potential successor to DRAM, SRAM and NOR Flash, ReRAM
(Resistive Random Access Memory), which might replace NOR flash and SRAM in the future,
FRAM (Ferroelectric RAM), and PCM (Phase-Change Memory). All these technologies are non-
volatile and in mass production but have yet to encroach on the established DRAM and NAND
flash markets.

Again, the problem is the lack of economies of scale, so today, these promising alternative memory
technologies are mostly confined to the embedded memory market such as RFID, mass-transit fare
cards, power meter readers, gaming systems as well as various other consumer, automotive and
industrial markets.

OTHER EMERGING SOLID STATE MEMORY TECHNOLOGIES

SITUATION ANALYSIS
Since the 1960s, semiconductors have been undergoing a constant pace of ~30% annual cost
reductions based on regular reductions in the size of the transistors on the chip. This process,
known as “scaling” involves reducing the length and width of transistors mainly through
lithographic techniques. The fact that this trend follows a constant slope was noticed in 1965 by
Intel founder Gordon Moore and has been given the name “Moore’s Law”. Gordon Moore’s 1965
paper noted that the number of transistors on a chip doubled every year or two, and, based on this
trend, Moore was able to project the transistor count of chips a decade into the future.

Memory chips have been following Moore’s Law since that time, doubling their density (the
number of bits on a chip) approximately every two years by continually shrinking the size of the
transistors on the chip. This doubling of transistor density each generation was achieved by
shrinking the lithography by approximately 30% each generation. Since area is linear dimension
squared, 0.70 * 0.70 = 0.49, or about half the size required for a given number of transistors for
each generation.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

If density scaling continues to follow Moore’s Law, both DRAM and 2D NAND flash were
expected to reach a point where they could no longer scale due to the fact that both of these
memories use charge storage to represent bits, and the charge that the memory cells could store
would eventually be too small to detect. 3D NAND flash has sidestepped this issue by increasing
layer count rather than reducing memory cell size, but DRAM scaling has slowed significantly.

At what point alternative memory technologies like MRAM, FeRAM, ReRAM, or others, might
be able to displace DRAM or NAND is unknown. Today, all of these alternative technologies are
still more costly per bit than DRAM or NAND flash, and this prevents them from being selected
as replacements except in those rare circumstances where the cost is less important than certain
important attributes they provide.

The alternative memory technologies are all “persistent” or “nonvolatile”, that is, they retain their
data even without power, and this allows them to be used as storage as well as working memory.
This is a key differentiator between NAND flash and DRAM – DRAM is fast (though not as fast
as SRAM) but it cannot store data when powered down (it is a volatile memory). NAND flash is
about one thousand times slower than DRAM, but retains its data without power. This allows
NAND flash to be used as storage while DRAM can only be used when powered.

The IEEE IRDS Road Map indicates that 3D NAND technology could extend the life of NAND
flash to 2030 and beyond, enabling chip densities to continue to increase even using process
lithography that would no longer be the most aggressive.

However, embedded NOR flash, used for code storage in embedded devices has reached a scaling
limit at 28nm and this has opened up opportunities to replace embedded NOR with persistent
memories that can scale to smaller size. MRAM and ReRAM are beginning to replace NOR flash
in some embedded products. Likewise, embedded SRAM may also face scaling limits at around
14nm and SRAM cells are very large because of the many transistors they use. For this reason,
there are some embedded devices that are starting to use MRAM or ReRAM to replace slower
SRAM caches in some embedded productsi.

FERROELECTRIC RANDOM ACCESS MEMORY (FRAM OR FERAM)

Ferroelectric memory, commonly known as FRAM or FeRAM, is a technology that has been in
the marketplace for over 30 years and one could argue that it has shipped in more products than
any other emerging memory technology because of its widespread use in embedded RFID
applications (such as fare cards). Discrete devices are available Fujitsu, Infineon, and Rohm, while
some Texas Instrument microcontrollers have embedded FeRAM along with ASICs from Fujitsu.
Densities range from 4kb to 16Mb.

Most FRAMs are used in mass-transit fare cards, RAID drives, gaming systems, and power meters,
where data must be rapidly written into nonvolatile storage while consuming very little power. In
RFID cards this power is generated from the interrogating radio waves. In other applications
power from a small charge on a capacitor is used to move data from a RAM into an FRAM in the

i
Emerging Memories Branch Out Report, Coughlin Associates and Objective Analysis, 2023

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

event of a power failure. All of these applications can justify FRAM’s higher cost structure
because of the extraordinarily low energy required to write data into the chip.

The “Ferroelectric” name is a misnomer given to Perovskite crystals that exhibit a state change
similar to the hysteresis curve of magnetic media. An atom located in the center of these crystals
can be pushed from one side of the crystal to the other by an electric current. The atom will stay
in that position until a current in the opposite direction moves it back. Current flow through the
device follows the familiar “S” curve of hysteretic devices, and can be used in its two stable states
to represent a logic “1” or a logic “0”. This storage mechanism is insensitive to radiation, unlike
the charge-storage mechanism used by DRAM and NAND flash, so there is some interest in its
use in avionics and military applications, although PCM and MRAM currently hold more appeal
than does FRAM in these markets.

A ferroelectric memory is manufactured by layering the Perovskite material onto standard silicon
CMOS logic as the dielectric of a storage capacitor. This adds cost to the wafer, making it
uncompetitive with established technologies. Ferroelectric materials disagree with silicon, so a
barrier layer must be used to protect the silicon from the ferroelectric material. This barrier layer
commonly consists of platinum, whose high cost does not increase materials costs appreciably
since insignificant quantities are required, but since platinum is very difficult to etch the cost of
processing a wafer is higher than that of standard silicon. As a result, and also because of the small
manufacturing volume for FRAM chips, FRAM wafers are more costly to produce than flash
wafers.

A further difficulty is that FRAMs have a larger two-transistor memory cell, and therefore a larger
die size, than single-transistor technologies like DRAM and flash memory, so their cost would be
higher even if the wafer processing costs were the same as those of established technologies.

In theory FRAMs’ inherent die size and processing cost disadvantages can be solved, but this
solution has not been put into practice. Today’s FRAMs cost significantly more to manufacture
than NAND flash memories of the same capacity. This indicates that it will be a long time before
FRAMs find use in mass-storage applications.

Fujitsu offers embedded FeRAM LSI products (RFID, authentication, and custom). Texas
Instruments offers an MSP430 microcontroller with FRAM (MSP430FR5xxx). Work by larger
manufacturing companies has brought much-needed R&D spending to the technology, without
which FRAM could never reach cost competitiveness with pure-silicon memories.

Applications that use FRAM find this technology attractive since it offers the high speed and low
power dissipation of SRAM with a cost structure that is theoretically similar to that of DRAM, and
is nonvolatile like flash. The blend of these three features means that the device could displace all
of today’s existing memory technologies should it reach the point where it can be manufactured
cost-effectively.

MAGNETORESISTIVE RAM (MRAM)

Magnetoresistive random access memory has also been in development since the mid-80s. Data
storage in these devices is in the form of storing magnetic spins vs electron charge storage utilized

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

in flash memory. Fundamentally, it is composed of a magnetic free layer separated from a

magnetic pinned layer by an insulating tunneling layer. The resistance of this magnetic tunnel
junction (MTJ) is determined by the orientation of the magnetic free layer with respect to the
pinned layer. If the two layers have the same magnetic alignment, the resistance through the
magnetic tunnel junction is minimized. If the two layers have opposite magnetic alignment, then
the resistance is maximized. The two types of MRAM currently in production are toggle MRAM
and spin torque transfer (STT-MRAM).

Other MRAM technologies are being developed with active research ongoing. Two types of
MRAM memory, Spin-orbit torque (SOT) and voltage controlled magnetic anisotropy (also known
as magnetoelectric RAM or MeRAM) are being considered as faster and lower energy MRAM
storage for future applications.

As a technology, it has been promoted as a potentially universal memory, replacing the

functionality of both DRAM and flash in one memory. Discrete devices are available from
Everspin, Honeywell, and Avalanche Technology. The main applications for stand-alone MRAM
to date have been avionics and military applications due to the insensitivity to radiation, fast write
speeds, and unlimited endurance. Currently, densities up to 1 Gb (STT-MRAM) are available.

MRAM is showing up in many embedded devices as a replacement for NOR flash for code storage
because of embedded NOR flash scaling limits and as a replacement for larger slower and higher-
level SRAM cache for applications such as AI inference devices. As the scaling used in embedded
devices decreases it is expected that the use of MRAM in embedded electronics will increase,
driving up production volume and lowering the costs of manufacturing.

RESISTIVE RAM (RRAM OR RERAM)

The resistive RAM or ReRAM is a large umbrella category that covers a number of subcategories,
namely PCM, PMCm, Oxygen Depletion Memory.

PHASE-CHANGE MEMORIES (PCM, PCRAM, OR PRAM): OVONIC UNIFIED MEMORY (OUM)

A number of companies are investigating phase-change memories, a type of ReRAM that uses the
crystalline/amorphous phases of a material to determine whether a bit is a “1” or a “0”. Most of
these companies were originally under license from Ovonyx, the company that invented and
owned the technology that it called the OUM or Ovonic Unified Memory, named for the material’s
inventor Stanford Ovshinsky. Intel, Micron, IBM, SK hynix, Samsung, and BAE are the most
visible, and many other licensees may be conducting undisclosed research.

In 2012 Ovonyx was liquidated, and many of the OUM patents were acquired by Micron
Technology.

PCM uses current to heat a small volume of a phase change material virtually identical to that used
for the storage layer in CD-RW disks. The heat changes the state of the material back and forth
between a crystalline and a polycrystalline structure, thus changing its resistance to a read current.
This change is reversible depending on the cooling profile. Today’s materials have been well
characterized but they are not fully understood. Research at the University of Pennsylvania points

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

to an alternate method of creating an amorphous area that could lead to improvements in both
programming methods and materials development.

Today’s PCM chips use chalcogenide glasses. Some chalcogenides (alloys containing one or more
group VI elements) exhibit reversible change between the disordered (amorphous) and ordered
(crystalline) atomic structure, the most common being Ge2Sb2Te5. When the crystal is amorphous
it exhibits a high electrical resistance, which lowers when the material crystallizes. The application
of highly localized heat drives rapid, reversible changes between the amorphous and crystalline
structure. A resistive heating element generates Joule heating, the temperature of which
determines whether the material comes to rest at an amorphous or a polycrystalline phase.

PCM has good technical specifications:

• Fast switching (<30ns),

• Good endurance (>1016 write/erase cycles),
• Imperviousness to cosmic radiation,
• Theoretical ease of integration into CMOS processes
• Scalability to smaller dimensions

PCM is an expensive form of memory today because the die size is larger than that of flash
implying higher costs, and wafer costs are high. The material promises to be able to migrate
beyond flash’s scaling limit, and this scalability is a compelling reason for researchers to consider
PCM as a replacement for NAND flash. PCM technology appears scalable to at least 5nm. Both
programming current and operating voltage decrease with process geometry, giving PCM an
advantage over flash. PCM should be relatively simple to add to existing CMOS or bipolar
structures. Stacked PCM structures promise to produce effective cell sizes smaller than those of
FRAM.

PCM commercial products have been introduced by Micron-Numonyx and BAE, and this
technology is one reason Micron gave for its 2010 acquisition of Numonyx. Samsung also
produced PCM for a time and used these chips in a low-end Samsung cell phone, but it appears
that no parts shipped outside of Samsung. Samsung is believed to have cancelled its PCM
development. The only company currently producing SoCs with embedded PCM is
STMicroelectronics for automotive applications.

In 2015, Intel and Micron announced a technology they call “3D XPoint Memory” (with the word
“XPoint” pronounced as “Crosspoint”). This was a form of PCM, although the companies just
said that it was a resistive memory. We will talk more about 3D XPoint Memory (which Intel sold
as Optane memory) later..

PROGRAMMABLE METALLIZATION CELL MEMORY (PMCM)

A good bit of research has been devoted to Programmable Metallization Cell Memory, or PMCm.
This ReRAM technology performs similarly to the chalcogenide glasses used in PCM but counts
upon metal threads migrating through the dielectric. These metal threads take the place of the
crystalline/amorphous areas in the PCM to provide either a high-resistance or a low resistance
path.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Adesto Technologies is shipping a product it calls “CBRAM” (Conductive Bridge RAM) which
is based on this sort of cell manufactured using chalcogenide glass with copper threads.

Another firm, Crossbar Technologies, promised to have a product by 2015, but has not yet shipped
or sampled a product. It has more recently been promoting use of this technology to create a
physically unclonable function (PUF) chip for product verification.

The benefits of these devices are similar to those of PCM, but where PCM is temperature sensitive,
requiring special solder flow processes to attach a pre-programmed PCM chip to a PC board,
PMCm is not temperature sensitive and can withstand the rigors of a standard soldering process.

At this early stage in the product’s life, it is too early to determine whether PMCm will achieve a
competitive cost advantage over other alternative technologies.

OXYGEN DEPLETION MEMORY

Recent strides have been made in the development of Oxygen Depletion Memory, another form
of ReRAM that employs a poorly-understood technique of moving oxygen atoms into and out of
a glass layer to increase or decrease that layer’s electrical resistance. The two companies that have
disclosed the most information on this technology are Unity Semiconductor, which was acquired
by Rambus, but later shuttered when the technology failed to perform as anticipated, and Hewlett
Packard, whose R&D labs made promises of the technology’s development in the early 2010’s,
but have missed every promised milestone to date.

The HP (now HPE) version of the technology has been dubbed “Memristor” and it was positioned
as a fourth missing basic electronic component after the resistor, capacitor, and inductor. It was
to have been the sole memory type used in HPE’s ambitious new computer architecture dubbed:
“The Machine”. Although HPE and partner SK hynix produced prototypes, insufficient progress
was made to continue to rely on the memristor, and the first rendition of The Machine was
converted away from memristor memory to DRAM use in the summer of 2015.

Oxygen depletion memory technology has had a number of near misses, but has never quite made
it to the sampling stage, so today it is farther behind than PCM, PMCm, MRAM, or FRAM.

COMPARING THE TECHNOLOGIES

This section will attempt to point out the comparative strengths and weaknesses of each
technology. As has been mentioned earlier, the most important factor is cost, and the memories
that account for the dominant share of the market are those with a lower cost structure than their
counterparts: DRAM and NAND flash.

NAND flash continues to be the cost leader in the cost-driven memory market by a very wide
margin, making it difficult to determine which of the emerging technologies is more likely to
supersede today’s memory technologies.
The fact that NAND is likely to maintain a decided lead for a number of years also complicates
matters. Since it will be a very long time before NAND flash memory is displaced by an emerging

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

technology, any one of these emerging technologies might achieve a breakthrough over the next
several years that might put it well ahead of the other emerging technologies.

Table 2 lists a number of salient features of today’s three leading memory types: DRAM, SRAM,
and NOR and NAND flash, and contrasts these against the attributes promised by MRAM, FRAM,
PCM and ReRAM: an umbrella category that groups together PMCm, and Oxygen Depletion
Memories.

Table 2. Solid State Memory Technology Comparisoni

Established Memory Types Emerging Memory Types
SRAM DRAM NOR NAND FRAM ReRAM Toggle STT PCM
Flash Flash MRAM
Nonvolatile? No No Yes Yes Yes Yes Yes Yes Yes
40-
Cell Size (f²) 40-500 6-10 10 0.03-5 10-32 4-50 16-32 4-50
160
Read Time (ns) 1-100 30 50 10,000 20-50 10-20 3-20 3-15 5-20
Write Time (ns) <1 1-10 105-7 104-6 50 101-5 1-10 1-100 106-9
Endurance ∞ ∞ 105 103-5 1015 103-9 ∞ 106-15 106-9
Rather
Write Energy Low Low High Med Low Low Low Low
High
Write Voltage None 2 6-8 12 2-3 1.2 3 1.5 1.5-3
Source: Objective Analysis & Coughlin Associates

Any of the four emerging technologies (FRAM, MRAM, PCM or ReRAM) will provide clear
technical advantages over the established memory types if it can be brought into price competition
with any of them. The price issue is paramount. Should any of the established technologies reach
a scaling limit one of the newer technologies is likely to take over its market.

Figure 8 shows another way of comparing many of the same attributes of these technologies. Note
that Ferro refers to FeRAM. This radar chart is set up so that each vector is either well satisfied
(100%) by a certain technology, or poorly satisfied (0%). The values represent a subjective rather
than an absolute valuation. The intent is to show that the new memory technologies (represented
by the dashed lines) have stronger technical attributes than today’s leading technologies, (solid
lines.) This advantage leads to the dashed lines’ running along the outside of the chart, rather than
dipping towards the center.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Read Speed
100%

Complexity Write Speed

50%

Scalable 0% Volatile

SRAM
Die Size Power DRAM
Flash
Ferro
Endurance
Magnetic
PCM

Figure 8. Radar Chart of Memory Technical Attributes

Source: Objective Analysis

The main advantage that any of these technologies has over flash is an ability to continue to scale
past the flash scaling limit. The line “Scalability” shows how each technology is expected to
follow standard semiconductor lithographic scaling. This scaling is the means that has been used
for the past half century to reduce the cost of semiconductors – the transistors are made smaller
every generation. Notice that scaling is limited for DRAM and NAND flash. If a technology can
be scaled but it does not have good cell density or array efficiency then it will still have trouble
displacing flash.

STORAGE CLASS MEMORY (SCM)

This section will focus on a category of non-volatile memory called Storage Class Memory (SCM).
This category is important since it set the stage for the 3D XPoint memory that Micron and Intel
introduced and then suspended development on. The experience with 3D XPoint (Intel’s Optane
memory) points out the difficulties of displacing or augmenting existing standalone memories.

SITUATION ANALYSIS
The sections above explained the need for emerging technologies to replace today’s entrenched
technologies once today’s technologies reach their scaling limit. We also noted that the emerging
technologies are all “nonvolatile”, that is, they retain their bits even without power, and this allows
them to be used as storage.

The ability to store data is a key differentiator between NAND flash and DRAM – DRAM is very
fast but loses its data when powered down, while NAND flash is one thousand times slower than
DRAM, but retains its data without power. This means that NAND flash can be used as storage
while DRAM can only be used as memory.

Emerging technologies fall somewhere between DRAM and NAND flash. They are almost as fast
as DRAM, and they retain their contents without power. This implies that the system’s memory,
which has been viewed as volatile storage since semiconductor memory replaced magnetic core
memory in the mid-1970s, could once again become nonvolatile or “persistent” as it was back in
the era of magnetic core memories. IBM gave these new memory technologies the blanket name
of “Storage Class Memory” or “SCM” since the company believes that new software architectures
will be developed to take advantage of the fact that memory will be persistent and can be used as
storage.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

With this conversion from DRAM and NAND flash to SCM will come significant changes in
computer architecture. Today’s software and CPU architecture has evolved since the mid-1970s
under the assumption that all memory is volatile and that all storage is many orders of magnitude
slower than memory. IBM’s SCM initiative was designed to anticipate upcoming changes in
which storage and memory would become the same, with the expectation that the underlying
assumptions of slow storage and fast, volatile memory would need to be reconsidered.

There have been numerous advancements designed to enable this change:

• The Storage Networking Industry Association (SNIA) established a nonvolatile memory

(NVM) programming model to streamline software stacks that communicate with SCM
• New memory modules (DIMMs) that pair DRAM with NAND flash to emulate nonvolatile
memory modules became a distinct market
• The Joint Electron Device Engineering Council (JEDEC, a standards-setting body for
semiconductors) defined naming conventions for three kinds of nonvolatile DIMMs:
NVDIMM-N, NVDIMM-F, and NVDIMM-P
• Intel and Micron announced a new SCM memory type they call 3D XPoint memory in the
summer of 2015 that was to be used as an SCM in future Intel processing platforms (to be
discussed in detail below)
• Intel introduced SSDs using its Optane memory (3D XPoint) as well as DIMM devices for
their processors
• Intel added new instructions to its standard Intel Architecture (IA) processors to support
SCM functionality.

THE INTEL/MICRON 3D XPOINT MEMORY

In the summer of 2015 Micron Technology and Intel Corporation announced a new memory type
called 3D XPoint (three D cross point). This was marketed as Optane (Intel) from 2017 to 2022.
Bit storage is based on a change in the bulk resistance rather than by charge storage as in the case
of flash. Some industry participants believe that this technology is a re-branding of the PCM
(phase change material) technology that these companies had been developing for years, but Intel
and Micron has said that the technology is different than PCM and uses chalcogenide materials
that are faster and more stable than traditional PCM materials.

3D XPoint could be considered to be a type of ReRAM (but so are traditional PCM devices). In
early 2021, Micron announced it would cease production of 3D XPoint and the fab in Lehi was
sold to Texas Instruments. In late July of 2022, Intel announced winding down the Optane
division.

As a technology, 3D XPoint fit in between NAND and DRAM. It was sold at approximately half
the price of DRAM, but about 5X the cost of NAND. Slower than DRAM, but faster than NAND,
there was a potential market niche since it was a fast non-volatile memory. It was positioned as a
new layer in the memory hierarchy – storage class memory.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

In that case, 3D XPoint could be considered to be a type of ReRAM. However, in 2021, Micron
announced it would cease production of 3D XPoint and the fab in Lehi, Utah was sold. In 2022,
Intel announced the winding down of the Optane division.

As a technology, 3D XPoint fit in between NAND and DRAM. It was approximately half the
price of DRAM, but about 5X the cost of NAND. Slower than DRAM, but faster than NAND,
there was a potential market niche since it was a fast non-volatile memory. It was positioned as a
new layer in the memory hierarchy – storage class memory.

BUSINESS/TECHNICAL ISSUES
The most important business problem standing in the way was that 3D XPoint memory had to be
sold as prices lower than DRAM. Achieving economies of scale is difficult when DRAM and
NAND are the two highest volume memories in production today.

As a result, it was likely that 3D XPoint memory needed to be sold below cost for an extended
period with the hope that demand will drive higher production volume and thus lower costs. If
this happened and the costs of production dropped below the prices the memory were sold for,
losses would turn into profits.

Intel had good reason to be willing to do this. By selling 3D XPoint memory at a loss the company
would be able to assure that the system’s speed keeps up with that of the processor, allowing Intel
to continue upgrading its customers to newer, faster, more expensive processors. If Intel can sell
a processor at a price that is $100 higher by losing $20 on the system’s memory then there would
be an overall profit. For other companies it is less clear why they would be willing to sell the
technology below cost to incubate the market. Indeed, Micron never manufactured any significant
3D Xpoint products for sale although it did announce some 3D XPoint products.

MANUFACTURING EQUIPMENT
The equipment used to manufacture this new technology was be identical to that used to produce
standard DRAM and NAND flash chips as well as other types of semiconductors. Micron
produced the first 3D XPoint samples in its Lehi, Utah USA NAND flash manufacturing plant
using a 20nm double-patterning process that is based on the company’s 20nm NAND flash
process.

MANUFACTURING PROCESSES
The structure of the 3D XPoint memory appears in Figure 9 below.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 9. Physical Mock-Up of a 2-Layer 3D XPoint Memory Structure

It appears that a lot of the processing that is used for the 3D XPoint memory is based on the basic
layered memory architecture that was pioneered by Matrix Semiconductor in the early 2000s prior
to SanDisk’s acquisition of that company. A microphotograph of the Matrix chip appears in
Figure 10.

Figure 10. Stacked Crosspoint Memory Array

Although these processes were new and are more difficult than standard semiconductor processes,
they were sufficiently well understood that they would not pose any undue difficulties. Today’s
PCM materials are based on germanium, antimony, tellurium, and aluminum.

GAPS AND SHOWSTOPPERS

The success of this effort depends upon Intel’s support of a new memory layer in its future platform
designs. Although Intel had a strong commitment to this effort and the size of Optane memory
cells was 1/10th that of DRAM several factors prevented this technology from every reaching
breakeven.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Optane SSDs never achieved the performance boost for users because of limitations in the SSD
interfaces for these products and so this market didn’t develop the way Intel had hoped. The
DIMM-based Optane products could only be used on Intel processors. Intel was a major driver of
the compute express link (CXL) memory technology, which could be used to allow pooling of
heterogenous types of memory with different performance characteristics, such as Optane
memory. An earlier introduction of CXL might have increased the demand for Optane memory.
Unfortunately, CXL did not come on the market until after Intel decided to discontinue
development of Optane.

The bottom line is that, although 3D XPoint has a smaller cell size than DRAM the wafer
production volume never reached the levels required for the production costs to drop below the
price that Optane had to sell for to be competitive to DRAM (about half the price of DRAM). This
inability to scale production for various reasons was the major reason why Optane memory did not
become the SCM it was hoped it would become.

It is estimated that Intel likely subsidized Optane memory production by over $7B between 2016
and 2022. After selling the company’s NAND memory business to SK hynix in 2020 and with
Micron’s withdrawal from making 3D XPoint memory in early 2021, Intel announced in late July
2022 that it would not continue to develop new Optane memory, although it would continue to
ship and support existing Optane memory products.

RECOMMENDATIONS ON PRIORITIES AND ALTERNATIVE TECHNOLOGIES

Although Optane memory was not able to achieve the volume it needed to become an established
player in the storage and memory hierarchy, its development spurred many other technologies that
benefit mass storage and non-volatile memory. In particular, it spurred development of PCIe-
based NVMe and CXL protocols and stimulated the development of software and firmware that
could optimally use fast non-volatile memory.

Other companies are trying to fill the gap in the storage and memory hierarchy left by the
withdrawal of Optane. These include NAND-based CXL devices from Samsung and SK hynix
(and likely other NAND flash manufacturers) that can either expand attached DRAM memory at
a lower price and introducing non-volatility or as part of a shared heterogeneous memory system
using CXL. In order to deal with the lower endurance of the higher-level NAND-flash based
products they are often using SLC flash to get closer to the high endurance (and performance) that
Optane would have offered.

Introducing standalone memories using new non-volatile technologies has big barriers due to the
need to get production volume up in order to bring prices down where the technology can be
competitive. However, introducing these memories in embedded products is much easier as long
at the new memory production processes can be easily combined with current CMOS processes.
For most of the emerging memories this is being done at the back end of the line, after CMOS
production is done.

For this reason, various emerging non-volatile memories are finding applications as memory in
embedded products serving many consumer, automotive and industrial applications. These
include replacing NOR flash for code storage and slower SRAM cache memory. Both of these

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

memories face scaling issues in embedded applications and SRAM cells are much larger than the
emerging memory cells and so more memory can be put in a given die area. Various companies
such as Fujitsu, STMicroelectronics and Texas Instruments are utilizing FeRAM and PCM in their
embedded devices.

All the major foundry companies, including TSMC, Global Foundries, Samsung and UMC are
offering MRAM or ReRAM embedded memories and several SoC devices have come on the
market, made in these foundries, for consumer, automotive and industrial applications. Growing
volume of embedded memory may help drive overall production volumes and thus decrease the
costs of using these non-volatile memories. This could create a virtuous cycle driving these
memories to higher volumes and making them attractive for new applications.

For this reason, academic and industrial laboratories continue to develop and refine new non-
volatile memories and industrial, automotive, consumer and aerospace applications will drive
adoption of these technologies in embedded products and for standalone applications as well.
Although 3D XPoint will not drive these changes, efforts on software and interfaces that can
support high performance and non-volatile memory that 3D XPoint inspired, will enable more
non-volatile memory applications.

Industry trade groups such as JEDEC, SNIA as well as professional technical standards
organizations, such as the IEEE, can play an important role in setting standards for these products
that will enable wide spread adoption and interoperability. Joint development and manufacturing
efforts such as the 3D XPoint efforts by Micron and Intel, could help drive the adoption of future
non-volatile memory technologies. Foundry support for these technologies will be critical for the
creation of niche markets for new non-volatile memories.

At this point, it is not clear what kind of memory could become that new Storage Class Memory,
but the system infrastructure is now being developed that will enable its usage – Compute Express
Link.

COMPUTE EXPRESS LINK (CXL)

CXL is a high speed serial link based on PCIe that enables high speed connectivity between CPUs
and devices (such as accelerators) and memory and is designed for high performance data centers.
It uses three protocols: PCIe-based block input/output protocol (CXL.io), a new cache-coherent
protocol for accessing system memory (CXL.cache), and device memory (CXL.mem). CXL was
initially developed by Intel, and then by the CXL Consortium which was established in March
2019 by founding members: Alibaba Group, Cisco Systems, Dell EMC, Meta, Google, Hewlett
Packard Enterprise (HPE), Huawei, Intel Corporation and Microsoft.

There were several competing consortia developing memory cache coherent standards prior to
CXL: OpenCAPI, Gen-Z, and CCIX. In February of 2022, Gen-Z specifications and assets were
transferred to CXL. On August 1, 2022, the OpenCAPI specifications and assets were transferred
to the Compute Express Link (CXL) Consortium. On August 3, 2023, the CCIX consortium
signed a letter of intent to transfer CCIX assets to CXL. So as of late 2023, the CXL consortium
has emerged as the industry focal point for coherent I/O. (www.ComputeExpressLink.org)

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

CXL 1.0 was released in March of 2019. In September of 2019, the CXL Consortium was
incorporated at CXL 1.1 was released. The CXL 2.0 spec was released in November of 2020, and
CXL 3.0 was released in August of 2022. CXL 1.0/1.1 allows a host CPU to access shared memory
on accelerator devices with a cache coherent protocol. CXL 2.0 adds support for CXL switching,
to allow connecting multiple CXL 1.x and 2.0 devices to a CXL 2.0 host processor, and/or pooling
each device to multiple host processors.

CXL 1.0/1/1/2.0 are based on PCIe 5.0 and supports x16, x8, and x4 widths natively. At 32GT/s,
a bandwidth of up to 64 GB/s can be achieved. CXL 3.0 supports PCIe 6.0 at 64 GT/s and up to
256 GB/s using a x16 link. It also expands CXL switches to multi-level, allows peer to peer direct
memory access, enhanced coherency and memory sharing. The Global Fabric Attached Memory
device is also a new feature in CXL 3.0 which enables disaggregated memory separate from hosts
(CPU, GPU, and other processing devices).

The three types of CXL devices are:

Type 1 (using CXL.io and CXL.cache) – specialized caching devices/accelerators (such as smart
NIC) with limited or no local memory. Such devices require coherent access to host CPU memory.
Type 2 (using CXL.io, CXL.cache and CXL.mem) – general-purpose accelerators (GPU, ASIC or
FPGA) with DDR, GDDR, or HBM locally attached memory.
Type 3 (using CXL.io and CXL.mem) – DRAM memory expansion boards and persistent memory.
Type 3 devices provide host CPUs with access to large amounts of DRAM or non-volatile memory.
Memory expansion beyond the directly attached DRAM via a CXL device will be possible with a
Type 3 device.

CXL is an open industry standard and an important enabling technology. The existence of three
prior consortia demonstrates the industry need. The standard offers a low latency and high
bandwidth connectivity standard between processors and devices such as memory, accelerators,
and smart I/O devices. It is expected to support the growing computational requirements for
machine learning, artificial intelligence, and high-performance compute.

If one were to summarize the promise of CXL in two words, it would be “endless memory”.
Memory on a CXL board could be shared between processors, or banks of memory could be
allocated to specific processors from the memory pool. While the initial CXL memory products
are for DRAM expansion, it is not clear that this is the best use case. DRAM on CXL increases
the latency and reduces the performance without directly reducing the cost. It offers to processors
and accelerators a big pool of slower DRAM. This might be useful if the DRAM capacity is the
main constraint in processing a workload. But a reasonably fast non-volatile storage class
memory, that is lower in cost than DRAM, would be a better fit here.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Hard Disk Drive (HDD) Technology

SITUATION ANALYSIS
Of current mass data storage technologies, in terms of storage capacity shipped, hard disk drives
(HDD) are by far the largest single component of the mass data storage industry. From its initial
introduction by IBM in 1956, this technology grew to play a vital role in computing data systems.
At this time the HDD market is undergoing a significant transition in markets with shrinking unit
volumes yet growing overall bit capacity due to a number of trends. These include the declining
desktop and notebook computer markets (signaling the demise of the 2.5-inch form factor in these
segments) as well as the nearly complete displacement of high-performance enterprise class HDDs
by SSDs (10k and 15K rpm drives have disappeared from the market leaving only 7200 RPM and
slower). These declines are somewhat offset by strong growth of HDDs, in unit volume and even
more in bit capacity, as warm/cool tier high capacity storage in data centers (nearline storage). It
is anticipated that the nearline market will stabilize the overall unit decline in HDD volume by the
last half of this decade as shown in Figure 11.

Figure 11. Shipped Disk Drive Volumes vs. Time, by Application

Source: Coughlin Associates

HDD’s are also undergoing a significant transition in recording technology. Perpendicular

Magnetic Recording technology has nearly reached its limit in areal density near 1.2Tbpsi. Higher
HDD capacities are currently being achieved mainly through putting more platters in the drive

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

enabled by the use of helium in place of air. Future areal density growth will depend on new
technologies such as heat assisted magnetic recording, three-dimensional magnetic recording, and
heated dot magnetic recording.

BUSINESS/TECHNICAL ISSUES
While HDD unit volumes have declined, with possible renewed growth later in the decade, prices
per GB have continued to decrease (although somewhat slower than in the past due to slower areal
density growth, Figure 12), driven by a number of key scientific and technical advances in
magnetic head, disk, interface, mechanical, signal processing, and manufacturing efficiencies.
While this has made HDD solutions suitable for a wider range of applications, particularly
consumer and entertainment applications, it has also created unrelenting pressure on the companies
that develop and manufacture HDD products to maintain profitability and business viability. More
stable HDD prices due to industry consolidation and better control of manufacturing inventories
have resulted in continued profitability by HDD companies, which could support the costs of
introducing the next wave of technology innovation.

Figure 12. Raw Storage Average Retail Price vs. Time

Source: Coughlin Associates

MANUFACTURING EQUIPMENT
The magnetic mass storage industry depends upon a complex set of manufacturing tools. The
critical lithography, pattern transfer and deposition equipment for head and disk production is
similar to that of the semiconductor industry, with similar critical dimensions, a unique materials
set and some different process requirements, such as extremely thin deposition layer thicknesses
(< 1nm), magnetic orientation, and 3-dimensional nanoscale topography. A unique aspect of HDD
manufacturing involves air bearing slider fabrication processes. HDD magnetic recording heads
require precise sub-micrometer lapping and polishing techniques for sub-nanometer control of
head-media spacing, magnetic head element and laser mount (for HAMR) geometry control.
Another unique feature of magnetic disk (as well as tape recording) is that a continuous surface
produced in a single manufacturing process is used for the recording media (unlike solid state and
optical storage devices).

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Mechanical assembly tools and test equipment are provided by a number of vendors, both North
American and international based. Manufacturers use a number of different providers of software
for manufacturing planning, process flow, inventory control, etc. The capital equipment required
for a single HDD manufacturing line can be in excess of $50-100 million. Thus, the decision on
commitment of additional production capacity must be made carefully. Key considerations here
include costs of manufacturing and manufacturing test equipment (such as drive burn-in and test),
and the need to reduce costs to meet customer requirements.

Storage industry associations such as the International Disk Drive Equipment and Materials
Association (IDEMA), Advanced Storage Research Consortium (ASRC), the Storage Networking
Industry Association (SNIA) and, in the past, the Information Storage Industry Consortium
(INSIC) give visibility to the equipment manufacturers, organize technical meetings on critical
issues, and form working committees to analyze and adopt common standards for component
specifications, manufacturing tests, storage management and product performance specifications.

MANUFACTURING PROCESSES
Development of key manufacturing processes can be carried out internally, providing proprietary
competitive advantage for drive vendors. Trade secrets for head and disk manufacturing and test
can be easily buried in manufacturing processes that are difficult to copy. However, these secrets
are sometimes not well enough understood, making process reproduction challenging. Figure 13
shows the general steps involved in manufacturing a magnetic recording head for a disk drive from
the wafer to the completed head gimbal assembly (HGA).

Figure 13. Hard Disk Drive Head Manufacturing Process

Source: Naniwa, I., Sato, K., Nakamura et al. Microsyst Technol 15, 1619–1627 (2009)

Measurements are critical in manufacturing processes. They are needed initially in process
development, and are vital in the transfer of information and components between vendors. This

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

has generated a need for improved process characterization and control, along with an
understanding at the engineering design stage of how a manufacturing process will operate in
volume production. Examples of process characterization and control in the storage industry are
the adoption of statistical tolerancing for mechanical components and assemblies, and the critical
role of supply chain and industry control software tools. These software tools lead to greater
efficiencies in supply chain management by automating the control of inventory and order
anticipation using the Internet.

The storage industry has seen some HDD companies adopt the use of independent production
contractors, which introduces an additional factor into the relationship between product design and
manufacturing. This includes the key role of manufacturing vendors in printed circuit board (PCB)
assembly and test, drive subassembly (head stack and actuator assembly), and hard disk drive
integration testing. In the past there were some drive companies that had contract manufacturers
build all or a substantial amount of their shipping drive volume. These drive companies could be
called fab-less drive companies. The industry has moved to almost all vertical integration, with
key head and disk technology development and manufacturing being brought increasingly in-
house.

For example, Western Digital Corporation acquired the assets of both Read-Rite and Komag
corporations for head and disk manufacture. HGST (Hitachi Global Storage Technology—
acquired by Western Digital) and Seagate Technology established internal head and disk
production capability. This integration allows close communication between the component
manufacturers and the HDD plant, to assure critical process problems are addressed early in the
manufacturing of a new product. The third remaining HDD company, Toshiba, buys all their heads
from outside suppliers, TDK and Showa Denko (now part of Resonac).

Up until a few years ago, the rapid advances in areal density of magnetic recording meant that
production lifetimes were shorter than in many industries. This required fast ramp-up to full
production, and made manufacturing process development a key constraint in new product
introduction. It also dictated that major changes from one product to the next must be more
evolutionary, to reduce risk associated with the new technologies. High manufacturing yields are
vital to assure a company’s ability to compete in this fast-paced industry. In the past, quite a few
technically strong companies have been forced to leave the industry because of their inability to
execute their yield ramps and cost controls on new product introduction.

MATERIALS
The leading companies are carrying out advanced materials development for their components and
the final products. Figure 14 and Figure 15 identify typical components in a magnetic hard disk
drive while Figure 16 shows the relationship of the head and disk indicating the head flying above
the disk and important characteristics of the head "slider."

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 14. Typical Hard Disk Drive with Key Components Identified (Photo of a Samsung Hard
Disk Drive)

Figure 15. Close-up of Actuator Positioning a Head Suspension (HGA) on a Disk (Photo of a WD
Drive)

Figure 16. Relationship of Head and Disk Including "Flying Height" of the "Slider"
Source: S. Wang and A. Taratorin, Magnetic Information Storage Technology, 1999

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Improvements in materials and processes have enabled several Japanese companies to become
suppliers of critical components and subassemblies to the magnetic recording storage industry.
Examples include disk substrates (high purity aluminum, as well as glass) and spindle motors (one
company in Japan (Nidec), supplies all disk drive companies with motors). The only independent
head vendor is a Japanese/Chinese partnership (TDK/SAE). TDK Headway/SAE has become well
positioned to offer volume production of state-of-the-art magneto-resistive heads. The only
remaining independent magnetic disk company is Showa Denko (now part of Resonac).

For the mass storage industry, critical advances are required in magnetic and non-magnetic
materials for functional head read and write performance at high areal densities and high
frequencies. These advances include: disk media and overcoats as well as heads capable of
supporting densities in excess of 160 Gb/cm2 (1,000 Gb/in2 or 1 Tb/in2); changes in media substrate
technology; and dielectric films less than 1 nm thick for advanced GMR (Giant Magneto-resistive)
and TMR (Tunneling Magneto-resistive) heads. In addition, technologies such as SMR (Shingled
Magnetic Recording), (TDMR) Two-Dimensional Magnetic Recording, HAMR (Heat Assisted
Magnetic Recording) an example of Energy Assisted Magnetic Recording and Heated Dot
Magnetic Recording (HDMR), will require HDD designs with new processes and materials.

HAMR technology requires an investment in optical materials and processing as well as new media
materials such as those based on FePt alloys. HAMR heads also require access to appropriate
lasers for the media heating.

Ongoing improvements in permanent magnet materials will be a critical enabler for more efficient
spindle motor and actuator design, and introduction of piezo-electric or MEMS (Micro-Electro-
Mechanical-Systems)-based actuators required new materials development. The first commercial
high-performance disk drives featuring dual stage actuators appeared on the market in 2005 in high
performance enterprise disk drives and this technology is now used in all hard disk drives. Figure
17 shows the basic structure of a modern TMR head transducer both as a schematic (16a) and a
TEM (Transmission Electron Microscope) photo showing the deposited layers (16b). Figure 18
shows an example of a piezo-electric-based micro-actuator for increasing the drive servo
bandwidth and thus the available track density of recording.

Starting in 2005 several disk drive companies began to ship perpendicular recording HDDs for
mobile computer applications. Figure 19(a) is an illustration of the older longitudinal recording
technology while Figure 19(b) shows perpendicular recording. The soft magnetic underlayer
(SUL) below the magnetic recording layer in the perpendicular media increases the effective
magnetic write field during the recording process allowing writing on higher coercivity media.

Perpendicular magnetic recording allows making thermally stable magnetic recording at higher
areal densities. For this reason, all drives in current production utilize perpendicular recording.
This transition was very rapid and has allowed areal densities to be pushed to near 1.2 Tbpsi.

Competitive advantages can accrue to drive and component manufacturers if they play a leading
role in the development of key material technologies since time to market for disk drive products
is so critical to a company's success.)

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 17. (a) Schematic of TMR head layers, including the Permanent Magnet (PM), Synthetic
Anti-Ferromagnetic (SAF), and Anti-Ferromagnet (AFM) Structures and (b) Air Bearing Surface
(ABS) view of a TMR head
Source: Seagate Technology

Figure 18. Example of a Piezo-Based Head Microactuator

Source: Hutchinson Technology Corp, now TDK

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

a) Longitudinal b) Perpendicular

Figure 19. Schematic Illustration of Magnetic Recording

Source: Seagate

ROADMAP OF KEY ATTRIBUTE NEEDS

Table 3 displays the current roadmap outlook of key attribute needs for Hard Disk Drive (HDD)
storage technology, including capacity and performance metrics, as well as key head, disk,
interface, and data channel characteristics.

Table 3. Magnetic Mass Data Storage Technology Roadmap – HDD

Unit 2022 2025 2028 2031 2034 2037

Industry Metrics
Form Factor inches 3.5, 2.5 3.5, 2.5 3.5 3.5 3.5 3.5
(dominant form factor
is bold)
Capacity TB 1-22 2-40 6-60 7-75 8-90 10-100
Market Size units 166 173 208 249 299 359
(M)
Cost/TB (avg.) $/TB 13.6 6.91 3.46 2.60 2.00 <2.00
Design/Performance
Areal Density Tb/in2 >1.0 >2.0 >4.0 >6.0 >8.0 >10.0
Rotational Latency ms 2-12 2-12 2-12 2-12 3-12 3-12
Seek Time* ms 3-5 3-5 3-5 2-5 1.5-5 1-4
RPM 4.2-10K 4.2-10K 4.2-7.2K 4.5-7.2K 4.5-7.2K 4.5-7.2K

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Data rate MB/se 140-600 140-600 180-1,200 200-1,200 210-1,800 220-1,800

c
Power watts 1.9-14 1.9-14 3-14 3-14 3-14 3-14
Key Component
Requirements
Read Head type TMR TMR TMR TMR TMR / CPP- CPP-GMR /
GMR LSV
Slider type & 5% 5% 5% 5% <5% <5%
size
(% of
micro,
3.86
mm3)

Disk type AlMg, Glass AlMg, AlMg, Glass, New Glass, New Glass, New
Glass, Glass, Substrate, Substrate, Substrate,
Disk Static Coercivity Oe 5,000-6,000 5,000-6,000 5,000-6,500 5,000-20,000 6,000-40,000 20,000-
HAMR HAMR HAMR HAMR 50,000
≥30,000 ≥30,000 ≥30,000 ≥30,000
Magnetic Recording Perpendicular Perpendicular Perpendicular Perpendicular Perpendicular Perpendicular
Technology TDMR, TDMR, , TDMR, TDMR, TDMR, TDMR,
ePMR, HAMR, SMR HAMR, SMR HAMR, SMR HAMR, SMR HAMR, SMR,
HAMR, Heated Dot
MAMR, SMR Recording

Electronics/Channel type LDPC LDPC LDPC LDPC Soft ECC, Soft ECC,
Iterative Iterative Iterative Iterative GPR TDMR TDMR
GPR GPR GPR (Turbo),
(Turbo), (Turbo) (Turbo), TDMR
Pattern TDMR
Dependent
Noise
Predictive
GPR
Channel Bandwidth MHz 500-2,000 500-2,000 500-2,000 500-2,500 >2,500 <3,000
SNR dB <20 <20 <20 <20 <18 <18
Actuator type Conventiona Conventiona Micro, + Micro, + Micro, + DSA Micro, + DSA
l/Micro, l/Micro, DSA DSA
+DSA +DSA
Spindle type Fluid Fluid Fluid Fluid Fluid Fluid
*Seek time is one third full stroke seek time and does not include micro-actuator local track or rotational latency.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

CRITICAL ISSUES
BUSINESS VITALITY
Ongoing vitality and success in the HDD industry is critically dependent upon companies’ ability
to sustain satisfactory profitability in a competitive environment of continuous global price
reduction and mounting pressure from alternative technologies. The ability to deal with eventual
price reduction will require constant improvements in development and operational efficiency,
while innovations in technology and manufacturing will allow leadership against competing
technologies to be maintained.

AREAL DENSITY GROWTH RATE

As mentioned in this report there are several promising technologies such as HAMR, HDMR
(Heated-Dot Magnetic Recording), TDMR, dual stage actuators, high sensitivity readers etc. that
look likely to enable higher areal densities. See Figure 20. However, these technologies are
expensive to develop and bring to market and may carry with them higher manufacturing costs
and production yield risks. There are now fewer players in the disk arena and less US government
funding for advanced technologies in this field. In a low margin industry such as HDD industry
has been, the lack of funding could seriously inhibit commercialization. It is thus likely that new
HDD technologies will continue to increase the annual areal density growth rate but it is unlikely
that the 50% or higher average area density growth rate seen in the past will return. ASRC’s areal
density roadmap suggests 20% CAGR from 2022 to 2035.

Figure 20. 2022 Areal Density Roadmap

Source: ASRC

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

COMPETITION FROM SOLID STATE DRIVES (SSDS)

SSDs are pressing HDDs in the area of cost at the low-end e.g. portable computer products, where
capacities are relatively modest and entry cost, form factors and power consumption favor
SSDs. At the high end, the superior performance (access time) of SSDs is gradually eliminating
the HDD high performance enterprise (server) business. The remaining market where there is
growth potential for HDDs is as high capacity nearline storage in data centers.

POWER CONSUMPTION
The power consumption of HDDs has become an increasing important issue as energy costs have
risen and customers have become more environmentally conscious. Some of the power
consumption of hard disk drives can be reduced by making the drives themselves more efficient
(better power supplies, more efficient electronics, use of helium rather than air, etc.) but a large
part of the problem is the excessive amount of time many drives spend in the idle mode i.e.
spinning but not transferring data. SNIA has worked with the EPA (Environmental Protective
Agency) on reducing the idle time in arrays and on developing an Energy Star rating for disk
drives. The sealed Helium drives, including many data center and enterprise HDDs enable over a
twenty percent reduction in idle power consumption.

SECURITY
In-disk drive encryption using FDE (Full Disk Encryption) is an excellent solution to the protection
of data on HDDs. It is being introduced in more types of HDDs with time and is gradually being
accepted by the marketplace. The HDD companies offer in-drive encryption in many of their hard
disk drives.

TECHNOLOGY NEEDS: RESEARCH, DEVELOPMENT, IMPLEMENTATION

For a given set of storage components (heads, media, spindles, etc.), writing a higher amount of
data to a given medium translates to an overall lower cost per gigabyte. Head and media
development work continues at corporate and university laboratories. Much of the advanced work
done on media and heads is basic science (where direct utilization of the information is less clear),
and support has to be found for this work. Because of business demand, companies manufacturing
magnetic storage products may not support long-range work needed to ensure future solutions in
all areas.

Universities are a natural place for the work to be conducted, and there have been seven centers at
major North American universities (Carnegie Mellon University, University of California San
Diego, University of California Berkeley, University of Minnesota, and University of Alabama, as
well as other groups world-wide such as those in Europe (Queen’s University of Belfast and the
University of York, in the UK) and Asia (Japan, Korea and Taiwan Universities as well as the
National Univ. of Singapore, and A-Star in Singapore) have contributed to advanced magnetic
recording development.

In addition, the guidance of industry-wide programs under the former IDEMA Advanced Storage
Technology Consortia (ASTC), and currently with the ASRC, have provided guidance for long
term “main line” storage programs for many years, and have fostered collaboration throughout the

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

industry. This collaboration was also enabled by metrology technologies developed by the
National Institute of Standards and Technology (NIST).

Research areas needing development to support continued growth of magnetic recording include
heated dot recording media, new reader designs, heat assisted magnetic recording (HAMR),
microwave assisted recording (MAMR), two-dimensional magnetic recording (TDMR), recording
channel and preamplifier design, multiple actuator servo systems, skew mitigation, and advanced
motor and multiple actuator design.

Disk drive track pitches are now < 100nm in size. Locating data and staying centered on such
narrow track widths during reading and writing, while the medium moves past the transducer at
almost 100 miles (160 km) per hour, is a huge engineering challenge. New actuator designs,
shingled write, heat assisted magnetic recording and heated dot magnetic recording (formerly
patterned bit media) technologies may offer solutions for further decreasing track pitch.

Taking advantage of advances in materials and processing, the disk drive industry is employing
aspects of micro-electromechanical systems (MEMS) in order to increase servo system bandwidth
and overcome motor bearing non-repeatable run-out and other eccentricities. This enables
narrower track widths and faster data transfer rates. Dual stage actuators in the head suspensions
using piezoelectric actuators built into the flexure are now used in all HDDs and many include a
third level of actuation in the suspension.

GAPS AND SHOWSTOPPERS

The rapidly increasing areal recording density in magnetic recording is creating a problem. As the
density of recording increases, for a given form factor recording device, the data that can be stored
on it increases. As the storage capacity on a surface increases, the time it takes to access a given
piece of data increases and the time to read or write all the data on the disk increases as well.
Although the average performance of disk drives to access data is increasing by about 10% per
year as shown in Figure 21, the disk access density (Access density = I/Os per second per gigabyte)
of disk drives is continually decreasing as disk drive capacity increases faster than disk drive
performance.
There are two main categories of challenge for HDDs

ACCESS DENSITY
While the areal density of HDDs has increased over the years, leading to devices with ever higher
total capacity, the inherent seek and rotational latencies associated with HDDs has created a related
problem that the I/O operations per second (IOPS) per byte (i.e., Access Density), is decreasing.
Given that the growth market segment for HDDs is migrating to high capacity nearline storage,
the Access Density problem is becoming increasingly acute, as shown in Figure 21. Lower Access
Density, in turn, translates into a $/TB problem.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 21. Disk Drive Access Density Is Decreasing

Source: Horison Information Strategies

A disk I/O is an input or output operation consisting of the seek time, latency time, and data transfer
time for a disk drive. As a consequence, it takes longer to get to a given piece of data and
especially, to move significant amounts of data. As the capacity of hard disk drives increases, the
potential for a larger number of concurrent users or applications goes up, resulting in longer queue
depths and slower response times. Late in 2017, Seagate announced multiple actuators to increase
drive IOPS. In its first generation, Seagate’s Multi Actuator technology uses two actuators on
same pivot which enables a potential doubling in IOPS while maintaining the same capacity.
Today both Seagate and Western Digital are shipping dual actuator disk drives and the number of
actuators on a given pivot could increase in the future, helping to keep the Access Density at
reasonable levels.

Another strategy to mitigate Access Density decrease is caching and multi-path storage
architectures. Hybrid HDDs (drives featuring flash memory for caching and buffering were
introduced in the marketplace in 2007. Also, some Solid State Drives (SSD) are marketed for
caching and application acceleration in combination with HDDs to provide a significant
performance boost while keeping the total $/GB costs at an acceptable level.

In May 2022 Western Digital introduced its OptiNAND SSDs. The 64GB of NAND in
OptiNAND host metadata, provides a non-volatile write cache (Armor-cache), enables high
resolution track interface management and enables enhanced servo capabilities. OptiNAND is
used in Western Digital’s 28TB UltraSMR HDD that offers 17% higher capacity than the non-
SMR version at 24TB.

SNR
The biggest technical hurdle to overcome in extending the areal density of magnetic recording for
HDDs is increasing the signal to noise ratio (SNR). It takes approximately 5-6 dB signal to noise
increase over the previous generation of HDD to double the density. This additional SNR can come
from the channel, the media or the heads. Better channels, improved reader and writer SNR and
improved media have all contributed to increasing areal density. There are still channel
improvements possible, but they are at the cost of increasing complexity (and thus more
processing) and we are approaching the theoretical limits of data decoding.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

TMR (Tunneling Magneto-Resistive) heads are in all modern HDDs. Write head improvements
in conjunction with improved media (smaller grains, higher coercivity, Hc) have been responsible
for significant gains in SNR. However, today’s write heads have reached the limits of known
materials in achieving write field strength with magnetic saturation fields of 2.4 Tesla.

The higher write field is necessary due to the increased magnetic anisotropy field, Hk, of the media
grains. In addition to increasing SNR, the grain size in the media must be decreased without
leading to thermal decay (thus requiring higher media coercivity). The introduction of HAMR and
other energy assisted recording technologies, as well as heated dot magnetic recording (HDMR,
formerly bit patterned media) offer solutions to extend density in the face of these increasingly
difficult constraints. With the introduction of HAMR, track pitch will be defined by the thermal
spot size and with HDMR, the patterned media bit spacing; therefore HAMR NFT (lithography,
etch and deposition) and patterned media process (assisted self-assembly) equipment will become
important as well.

In late 2023 Seagate began volume shipments of 32 TB HDDs with 10 disks using heat assisted
magnetic recording (HAMR). Seagate projects higher areal density drives using HAMR in the next
few years. Seagate plans for 50TB+ HDDs with areal density over 3.0 TBpsi by 2025, following
more than a 20% annual increase in areal density.

Disk drive track density is a key technology to increasing areal density, particularly since linear
density may be limited by thermal decay sensitivity. The track pitch is defined by the width of the
write (and read) heads. These dimensions are created using lithographic techniques borrowed from
the semiconductor industry.

RECOMMENDATIONS ON PRIORITIES AND ALTERNATIVE

TECHNOLOGIES
Key to the continued health of the magnetic mass storage industry is innovation in a wide variety
of technologies. These include: head magnetic properties for better writing performance and
signal-to-noise ratios on read-back, magnetic media to support higher areal density, improved
materials, advanced interface and air bearing designs, precision mechanical components,
contamination control, high speed electronics, and signal processing technologies. Drive sizes less
than 1.8” have disappeared-likely forever. 2.5-inch drives are still in production, but their number
will continue to fall as sales of HDDs into PCs slows with growing adoption of SSDs. 3.5-inch
HDD, particularly high-capacity drives for nearline applications will be the remaining HDD form
factor in a few years.

SHINGLED MAGNETIC RECORDING (SMR)

One approach to improving areal density in hard disk drives is shingled writing. With few changes
in existing heads and disks this technology has shown a practical increase in areal density of 10-
18 percent (Western Digital’s UltraSMR with OptiNAND has 18% higher capacity that the
conventional magnetic recording or CMR version). Shingled recording (Figure 22) uses a
recording head with a stronger, but asymmetric write field. This head is utilized with sequential
track writing and a track pitch that is significantly smaller than the head effective magnetic write

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

field width. The overlapping write process leaves behind a written track, which while much
narrower than the write head width, is easily read by the even narrower TMR read head.

Figure 22. Shingled Recording and Head Magnetics

Source: Hitachi GST

Note that shingled writing has been used for years in magnetic tape recording. Shingled recording
was introduced in LTO-2 tape products in 2003 (nearly a decade before the introduction of
shingled recording in HDDs in 2013.

While this technology allows random reads, it does not readily accommodate random writes. Due
to the nature of the write process, a number of tracks adjacent to that being written are overwritten
or erased, in whole or in part, in the direction of the shingling progress, creating so-called “zones”
on the media, which behave somewhat analogously to erase blocks in NAND flash. This implies
that some special areas on the media must be maintained for each recording zone, or group of
zones to allow random write operation, or random writes must be cached in a separate non-volatile
memory.

Further, the writing of sequential data must be maximized to minimize the writing of random data.
To accomplish this, special algorithms, caching, and metadata (data which describes data) are
needed in a shingled HDD architecture. In 2022 Seagate said that nearly 25% of their nearline
drives had been SMR and SMR has also been widely used in external backup HDDs. WDC has
made similar statements.

SMR operation is most suited for sequential write workloads and has the potential to negatively
impact performance for random write-intensive applications. To better manage these impacts in
enterprise systems, particularly increased write latency, standards were developed which allow the
host system to manage the SMR device’s write characteristics.

SEALED HELIUM DRIVE

The density of helium is one-seventh that of air, enabling advantages such as less drag force acting
on the spinning disk stack, thus substantially reducing the required spindle motor power and the
fluid flow forces buffeting the disks and the arms are substantially reduced allowing for disks to
be placed closer together and to place data tracks closer together (see Figure 23). The lower shear
forces and more efficient thermal conduction of helium also mean the drive runs cooler and emit

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

less acoustic noise. Finally, the use of helium also allows the use of larger diameter disks, thereby
increasing total drive capacities.

Figure 23. Western Digital 22TB and Seagate - 20TB Sealed Helium Drives
Shipment of production models of these drives commenced in 2013 but market growth was slow
at first until production volumes improved. Seagate Technology and Western Digital began
shipping similar products in 2015. Both vendors are now shipping 20+TB models and these
products are popular for data center and enterprise applications due their higher energy
efficiency and enabling more disks and heads per drive.

HEAT ASSISTED MAGNETIC RECORDING (HAMR)

Perpendicular recording is used in all commercial disk drive products. Areal Densities have been
demonstrated using perpendicular recording at 2 Tb/in2 . In spite of this progress, serious
challenges face continued use of pure perpendicular recording without some enhancement.
Conventional perpendicular recording or so-called conventional magnetic recording (CMR) is not
extendable much beyond 1 Tb/in.2. CMR density is constrained by (a) the need to reduce media
grain size for higher areal density (AD), (b) the requirement to increase magnetic anisotropy to
maintain thermal media stability and (c) the physical limits of write fields to record on the media.
These constraints are commonly referred to as the Trilemma.

Heat Assisted Magnetic Recording (HAMR) is an alternative to increase magnetic recording areal
densities by heating during the recording process. HAMR uses the fact that the coercivity (HC)
drops continuously, to practically zero, when a magnetic material is heated near its magnetic
ordering temperature or Curie temperature (TC). The HAMR writing process uses this effect by
heating the media to an elevated temperature where HC is below the writing field of the head. The
heated medium region is then rapidly cooled down to ambient temperature after the head field is
applied to write on the medium. This permits writing of media with much higher room temperature
HC and with smaller grains than conventional perpendicular recording, resulting in much higher
areal density. A sketch illustrating the HAMR writing process is shown in Figure 24.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 24. Curie Point Writing Using Head Assisted Magnetic Recording
Source: Seagate Corporation

A HAMR recording system needs the additional capability to heat extremely small spots on the
medium along with all the features of CMR. Seagate’s shipping HAMR system uses laser light
focused on a Near Field Transducer (NFT) as heat sourcei. This system couples laser light into a
specially shaped waveguide for light delivery to a NFT to form a deep submicron thermal spot on
the media. A schematic overview of a proposed HAMR recording system is shown in Figure 25,
to introduce its most important components. Figure 26 shows an approach where the laser is built
into the head itself.

Both the HAMR head and media in these two approaches can be manufactured on existing head
and media manufacturing lines with some modification to existing processes and with additional
components (such as lasers) and assembly tooling. Seagate has been shipping HAMR drives for
data center and enterprise evaluation over the last couple of years and plans to ship these drives in
volume starting in 2023.

Figure 25. Sketch of basic proposed HAMR recording showing head design w/light impinging on
the grating
Source: Seagate Technologies

i
Rottmayer et al, “Heat Assisted Magnetic Recording”, IEEE Trans. Mag., Vol. 42, No.10, p 2417, 2006

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 26. HAMR head with the laser source built into the head
Source: Seagate Technologies

ENERGY ASSISTED PERPENDICULAR MAGNETIC RECORDING (EPMR)

ePMR is a technology currently being used by Western Digital to increase the capacity of their
HDDs. ePMR increases BPI by applying an electrical current to the main pole of the write head
throughout the write operation. This bias current enables more consistent, and faster switching of
the write head, thus reducing timing jitter. Higher BPI is achieved when individual bits of data
can be written closer together, which leads to higher areal density. Western Digital discovered
this effect and is currently the prime user of this technology.

MICROWAVE ASSISTED MAGNETIC RECORDING (MAMR)

Microwave Assisted Magnetic Recording (MAMR) is an alternate Energy Assisted Magnetic
Recording concept to extend AD. MAMR seeks to improve writability by supplementing the
perpendicular write field with a localized RF field which is typically supplied by a small Spin
Torque Oscillator (Figure 27).

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 27. MAMR concept

Source: Mallory et al, IEEE Trans Magn, 50, 3001008, 2014

Modeling has suggested MAMR may extend AD to near 4 Tbpsi if STO requirements can be
achieved and suitable media engineered. (Ref. J Zhu, IEEE Trans Mag 40, 3200809, 2014). To
date very limited experimental AD gains have been publicly demonstrated. A version of MAMR
(Flux Control-MAMR) is being used by Toshiba in some current high-capacity HDDs and the
company has shown a path to using Microwave Assisted Switching MAMR (MAS-MAMR) for
future HDD capacity gains. (Figure 28)

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 28. Toshiba MAMR/HAMR Roadmap derived by Chris Mellor in The Register Jan. 13, 2022

TWO DIMENSIONAL MAGNETIC RECORDING (TDMR)

Today’s granular media is composed of grains with no directional information on the media. See
Figure 29. Thus, the surface of the magnetic medium is a two-dimensional environment.
Conventional magnetic recording defines which direction is along-track and which direction is
cross-track.

Figure 29. Media surface is a two dimensional environment

Source: Seagate Technology

The TDMR concept utilizes the two-dimensional magnetic surface for higher Areal Density (AD)
and potentially better reading performance. Multiple generations of TDMR are envisioned. (Ref:
TDMR Roadmap, F1, TMRC 2015) with progressively growing read head complexity to facilitate
more sophisticated data encoding and density. Specifically, it is based on:

• Reading multiple adjacent tracks either by multiple read elements (See Figure 30) or multiple
spins, and
• Processing those tracks jointly.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 30. Dual reader TDMR concept.

Source: IDEMA ASRC

TDMR eliminates the direct relationship between the track pitch and read-head cross-track profile;
it allows Inter-Track-Interference (ITI) between tracks which can be resolved by processing
adjacent tracks jointly. This enables different Linear Density (LD) and Track Density (TD) points
for the overall system to yield higher AD and possibly better read performance. Initial TDMR
products were introduced by Seagate in 2017. The technology is currently used on many high-
capacity HDD models.

HEATED DOT MEDIA

The combination of HAMR and patterned media (Heated Dot Magnetic Recording – HDMR) will
be required to bring areal densities of magnetic recording to 10 Tb/in2. Patterned media has
discrete elements of the magnetic material distributed in an orderly fashion across the disk surface.
These patterned dots become the magnetic bits when the head writes on them. Patterned media
will be combined with HAMR recording to create 10 Tb/in2 magnetic recording areal density
within the next 15 years. Figure 31 compares conventional magnetic recording media to patterned
media.

In patterned media, the servo information is defined by lithography. While this would potentially
save equipment costs for servo writers, the positioning accuracy of lithographically defined servo
information needs improvement to be useful. These risks and in particular the perceived capital
expenses for moving to patterned media have convinced the HDD industry to deal with patterned
media as advanced research with introduction into volume manufacturing more than 10 years out.
Figure 32 shows some examples of patterned media including servo and address patterns.

In patterned media the nano-scale track and bit dimensions would be defined only by lithography.
There are several technical candidates for this enhanced lithography.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 31. Comparison of conventional magnetic media to patterned media

Source: Hitachi Global Storage Technologies

Figure 32. SEM images of self-assembled patterned magnetic recording media showing servo and
address patternsii

The likely approach for creating the physical media patterns is the use of block co-polymer self-
assembly and pattern multiplication (i.e., SADP, SAQP, SAOP) to reduce the pitch such as used
in semiconductor. The use of Di-Block Co-polymer (DBCP) self-assembly can be used to define
a patterned media. This pattern must then be transferred to form the patterned media. This method
has produced bit densities of 5Tbpsi with a pitch near 15nm. DBCP technology is the result of
certain physical and chemical properties of unique polymeric substances.

ii
Yoshiyuki Kamata, Akira Kikitsu, Naoko Kihara, Seiji Morita, Kaori Kimura, and Haruhiko Izumi, “Fabrication
of Ridge-and-Groove Servo Pattern Consisting of Self-Assembled Dots for 2.5 Tb/in2 Bit Patterned Media,” IEEE
Trans. Mag., 47, 51-54 (2011)

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

These have the properties of self-assembly with highly consistent pitch which can be combined
with a long-range guide pattern such as a simple, low-density e-beam generated pre-pattern. This
minimizes e-beam lithographic write times since the patterns produced function as alignment
marks to assist the ordering of the self-assembly process. In addition, the technique could be used
to write pre-patterned imprint masters and consequently could reduce the costs of making e-beam
masters.

Figure 33 indicates how block copolymers can be used to create these patterns. To increase SNR
it is further desirable for the media bits to be rectangular rather than round. WDC and Seagate
have overlaid printing of concentric rings with radial lines to produce rectangular bits. Recording
demonstration to 1.6 Tbpsi have been demonstrated by Seagate.

Figure 33. Block co-polymer self-assembly

Source: Photo Courtesy of IBM Almaden Research Center

The use of e-beam exposure, today’s process for fabrication precision optical masks, can be
considered the ultimate in lithographic processing, although as previously indicated today’s e-
beam tooling and resist formulations result in very low throughputs. It is likely that e-beam
lithography will be used to create master patterns that can be used to make working imprint devices
for creating imprinted patterns on magnetic surfaces. The process of creating working imprint
surfaces from e-beam defined masters could involve either two or three generations of working

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

stampers. One e-beam lithography generated master can be used to create several generations of
imprinting surfaces, each of which may be capable of imprinting 100,000 disks.

This procedure is feasible since the e-beam technology required for nanoimprint master exposure
could require many hours of production time, which can be amortized across a great many disks.

Patterned media requires magnetic bits to be located by lithography, whereas today’s magnetic
media accomplishes bit locations along a track through prewritten servo patterns and the actual
write process from the magnetic head. Defects originating from lithographic errors in either
exposure (from masks defects) or during the etching process include shorting of adjacent bits
(bridging), missing bits, and diameter or positional variations. The latter would result in timing
errors and reduced magnetic fields. These issues must be resolved to make this technology ready
for production use.

Patterned media can be made by a variety of processes from “self-organizing” structures that could
include lithographic and nanoimprinting techniques. Figure 34 shows how patterned media may
be manufactured.

Figure 34. Possible patterned media production process

Source: Intevac

FULL DISK ENCRYPTING DRIVES

Full disk encrypting (FDE) drives encrypt everything that is written to the drive and decrypt
everything that is read from the drive. So, encrypted data exists only while “at rest”, while stored
on the drive. The purpose of FDE is to provide the following:

• Protection from loss or theft of the drive (or the computer containing the drive): Since the
data is encrypted, access by unauthorized users is not possible. All 50 states, the District
of Columbia, Guam, Puerto Rico and the Virgin Islands have laws requiring private

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

businesses, and in most states, governmental entities as well, to notify individuals of

security breaches of information involving personally identifiable information, UNLESS
the data is encrypted.
• Re-purposing or end-of-life: Under privileged (administrator) control, the cryptographic
key can be erased, rendering the encrypted data inaccessible, thus “sanitizing” the drive,
which can then be used as a re-purposed FDE drive (new key) or disposed.
• “Rapid erase”: key erasure is nearly instantaneous, so that ‘sanitization’ does not take
hours that traditional methods require, such as data overwriting.

The cryptographic function is implemented in dedicated hardware (circuitry) directly on the drive,
so that system performance is not impacted. The cryptographic key never leaves the drive,
eliminating one layer of key management required for software-based solutions. Instead,
authentication keys are used to unlock the drive and invoke the FDE function.

Initially, FDE drives were available for portable devices, like laptops, due to the ever-increasing
mobility of the workforce and the high frequency of laptop loss and theft. However, the data center
is also subject to hard drive “mobility”, not only misappropriation of the drives, but leaving the
data center for maintenance and end-of-life. For this reason, FDE is now available on all new
design enterprise and data center class drives. In addition, having the cryptographic function
directly on the drive (instead of somewhere else “upstream” in the storage system), simplifies data
center planning and configuration management, supports expansion with no performance impact,
and eliminates the need for data classification.

FDE in hardware, directly on an HDD, has been standardized by the Storage Workgroup of the
Trusted Computing Group (TCG). A subgroup of TCG has spelled out the details of
(authentication) key management for enterprise drives and multiple drive configurations within an
enterprise. The TCG Storage Specification defines an architecture for a variety of security
functions built directly into storage devices, including FDE. (download available at:
https://www.trustedcomputinggroup.org)

The Institute of Electrical and Electronic Engineers promulgated a comprehensive data sanitization
standard covering data security in 2022, IEEE 2883-2022, that recognizes FDE as cryptographic
erase, a medium security methodology.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Magnetic Tape Storage

SITUATION ANALYSIS
Linear tape technology uses the same basic magnetic recording principles as hard disk drives
(HDDs) and leverages many of the technologies developed by the higher volume and more
advanced HDD industry. For example, linear tape drives first adopted AMR, then GMR and most
recently TMR reader technologies that were first developed for use in hard drives up to 10 years
before being used in tape drives. The latest enterprise class tape drive, the IBMi TS1170, which
was released in Aug. 2023, operates with a native cartridge capacity of 50TB. Although state of
the art tape drive bit length is only about 3.5x larger than that of the highest capacity disk drives,
tape drive track width is about 12x larger than disk and hence tape areal density is about 42x lower
than HDD.

However, the latest generation of tape cartridges contain more than 1000m of 12.65mm wide tape,
e.g., the TS1170 user tape length is more than 1200m, resulting in a useable recording surface area
that is about 100 times larger than that available in a state-of-the-art high-capacity HDD with ten
platters that are > 95mm in diameter. Hence the capacities of the latest generation enterprise tape
format is roughly double the highest capacity HDDs on the market at the time of its release (i.e.
22TB CMR and 26TB SMR). The fact that tape systems operate at a roughly 42x lower areal
density than state of the art HDD implies that tape systems have the potential to continue scaling
areal density and capacity for multiple future generations before having to face the challenges
resulting from the super paramagnetic effect (i.e., the magnetic recording trilemma) that the HDD
industry is currently struggling with.

Tape drive developers have adopted a variety of technologies originally developed for disk drives.
These include head technologies and data channel algorithms such as NPML (Noise Predictive
Maximum Likelihood detection). The challenge for tape drive developers is to adapt these
technologies for use in tape drives and to develop additional technologies to deal with the unique
challenges that arise from parallel recording on a flexible tape media that runs in contact with the
head. Rather than a single active recording channel as in a disk drive, state of the art tape drives
record and simultaneously read verify 32 tracks in parallel.

In recent generations of IBM Enterprise and LTOii drives, this is achieved with a three-module
head design composed of a read module with 32 data readers and two servo readers sandwiched
between two data writer modules which each contain 32 data writers and two servo readers, as
illustrated in Figure 35. In this architecture, the left writer module is used for writing data in the
forward tape direction and the right writer module is used for writing in the reverse direction and
the center reader module is used for data read back and read-while-write verification in both
directions. In addition to the reliability provided by read-while-write verification, tape drives also
implement powerful error correction codes (ECC) to achieve bit error rates that are four orders of
magnitude better than disk products.

i
IBM is a trademark of International Business Machines Corporation registered in many jurisdictions worldwide.
ii
LTO is the trademark of HP, IBM, and Quantum in the Unites States and other countries.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 35. (a) Photograph of a 3 module, 32 channel tape head, (b) Illustration of a 3 module 32
channel tape head. The center reader module (CR) contains 32 data readers (R01-R32) and 2 servo
readers (S1 and S2), and the left (LW) and right writer (RW) modules each contain 32 writers (W01-
W32) and two servo readers (S1 and S2). The read transducers in the reader module are aligned
with the write transducers in the writer modules to enable read while write verification.

The multi-channel nature of modern tape drives necessitates the design of a set of custom ASICS
for driving all the parallel channels, i.e., driving the writers, pre-amplification and filtering for the
readers and to implement the digital data and servo channels as well as the ECC algorithms.
Compared to HDD, the single channel data rate of a tape drive is much lower, which simplifies
some aspects of the ASIC designs, however support for many parallel channels creates additional
complexity.

Moreover, the removeable nature of tape media and requirements for interchangeability and
backwards compatibility as well as the use of contact recording in which debris can temporarily
be deposited on the head, all necessitate that the data and servo channels of tape drives are designed
to be much more adaptable than those implemented in HDD. In addition, the support of a wide
range of tape speeds required for the variable data rates provided by tape drives in order to match
the host’s transfer rate imposes additional ASIC design challenges.

The continued scaling of ASIC technology to smaller transistor sizes has enabled tape drive
designers to develop new ASICs with the capability to process the increasing number of parallel
write and read transducers in the head without significant increases in the ASIC die size.
Compared to HDD, the number of tape drives shipped per year is orders of magnitude lower and
hence the number of ASICS required is also much lower. This challenge of low ASIC volumes is
compounded by the rising costs of designing custom ASICs as the technology nodes scale to
smaller transistor sizes and more expensive lithography and mask technology. The tape industry
therefore amortizes the cost of designing and manufacturing these ASICs by reusing ASICs in
multiple generations and across different product families.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

The most recent enterprise and LTO tape formats, IBM TS1170 and LTO-9, both support the built-
in file system called the Linear Tape File System (LTFS) that was first introduced with the LTO-
5 tape format. LTFS expands the capabilities of tape drives and tape library systems to allow
Network Attached Storage (NAS) like single tape behavior and in library systems allows faster
access time to data. RESTful APIs using LTFS tapes have also enabled object-based storage built
around the use of magnetic tape.

BUSINESS/TECHNICAL ISSUES
Historically, tape capacity has been scaled primarily by areal density scaling with additional
capacity gains achieved through improvements in format efficiency and reducing the tape
thickness to increase tape length. As with HDD, tape media advances are key to continued tape
scaling. As areal density is increased, improvements in media are required to compensate for the
loss in signal and signal to noise ratio (SNR) resulting from the smaller bit size and reductions in
the thickness of the magnetic recording layer that help to enable increases in linear density.

Tape media recording performance improvements are typically achieved through a combination
of technologies that include reducing the size of the magnetic particles in the recording layer to
reduce media noise, developing new particles with improved magnetic properties, improving the
dispersion of the magnetic particles, orienting the particles during the coating process, reducing
variations in the thickness of the recording layer and reducing the media roughness in order to
improve magnetic spacing.

Unfortunately, making the media smoother tends to increase tape-head friction which can degrade
recording performance and reduce the ‘runability’ of the tape. These effects can be minimized
through careful engineering of the tape surface roughness, improvements in lubrication technology
and optimization of the geometry and topography of the tape head to minimize contract area.
Additional tape path components and features will also likely be adopted to deal with increases in
startup friction/stiction and tribology issues during tape transport that arise from smoother media.

In state-of-the-art linear tape drives, 32 equally spaced data tracks are written and read back in
parallel in one of four data bands that are about 2.7 mm wide as illustrated in Figure 36. The
written track width is much smaller than the spacing between adjacent writers and hence many
passes back and forth along the length of tape are required to fill a data band. The roughly 2.6 mm
span of the array of transducers in combination with tape dimensional stability (TDS), i.e., changes
in width of the thin flexible tape, makes it challenging to keep all the writers aligned with the
desired track locations during write operations and all the readers centered over the written tracks
during readback.

The dimensional stability of the tape depends on thermal expansion, hygroscopic expansion,
operating tension and stress induced creep during storage in the cartridge reel. A user can record a
tape at one extreme of environmental conditions and will expect to be able to recover the data at a
different environmental extreme, perhaps years later. Historically, TDS was dealt with by
including a component in the track width budget to account for the changes in tape width due to
TDS and therefore ensure that the readers don’t get too close to the track edges during readback.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

The other major component of the tracking budget accounts for track following errors, i.e., errors
in positioning the head relative to the moving tape during read and write operations.

Reductions in track width with each generation of tape drives have therefore necessitated
improvements in TDS and improvements in track following accuracy to ensure reliable operation.
Recently, tape drives have implemented active TDS compensation to ensure that all 32 read and
write transducers are placed at the correct location on tape in the presence of changes in tape width
due to TDS effects. This innovation relaxes somewhat the need for continual improvements in tape
dimensional stability which were becoming increasingly difficult to achieve. Hence, continued
track density scaling will require continued incremental improvements in active TDS
compensation and improvements in track following fidelity.

Figure 36. Illustration of the four data band / five servo band tape layout used in the LTO format.
LTO generations 7 to 9 use a 32-channel format illustrated on the right side of the figure in which
each data band is subdivided into 32 sub-data bands. 32 tracks are written in parallel with forward
wraps (tracks) written in the upper part of each sub-data band and reverse wraps written in the
lower part. Multiple passes back and forth along the length of tape are required to fill each data
band.

The current tape market is dominated by the Linear Tape Open (LTO) format with a smaller share
held by the IBM TS11xx enterprise format. Current LTO media manufacturers include Sony as
well as Fujifilm who also manufactures media for IBM’s TS11xx tape drives. HPE, IBM and

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Quantum all provide TPC (Technology Provider Companies) certified LTO drives. Multiple
companies including Dell, Fujitsu, HPE, IBM, Quantum, Spectra Logic, Tandberg, etc., offer tape
library solutions.

Tape systems provide a very low total cost of ownership for storing large volumes of data. For
example, a recent 10-year TCO (total cost of ownership) study from ESG found an 86% cost
reduction for an LTO8 tape solution over an HDD based solution [1]. In addition, tape systems can
achieve very high streaming data rates but with the penalty of access latencies in the range of tens
of seconds. Hence, tape systems are particularly well suited for archival data storage applications,
i.e., the long-term preservation of infrequently accessed data. The built-in physical air gap
provided by the removeable nature of tape cartridges provides an additional layer of protection
against cyber-crime or accidental data deletion.

The removeable nature of tape also results in a very low power consumption per PB of capacity
that contributes to both the low TCO and the very low CO2e footprint of tape systems. A recent
IBM study compared the 10-year CO2e footprint of a 27 PB tape archive to an open compute
project (OCP) Bryce Canyon HDD solution and found a 96% lower CO2e for the tape solution
and more than 90% less power consumption. The combination of these benefits has been driving
growth in the tape market with particularly strong growth amongst hyperscale cloud companies.
In 2021 the overall tape market grew by about 14%, with approximately 2/3 of that growth driven
by the Hyperscale market segment [2]. The tape market is projected to continue growing with a
projected growth of 64% by 2025.

Figure 37 shows an LTO-9 (Linear Tape Open Generation 9) full-height tape drive and cartridge.
Tape drives and cartridges are often used in automated library systems. Figure 38 (a) and (b)
shows the IBM and the Spectra Logic Tape Library offerings, respectively. Not all of the library
types shown in Figure 38 are designed for use with enterprise drives and at the time of writing,
only the TS4500 provides support for the very recently introduced TS1170 drive. Therefore, to
facilitate comparisons between different library types, the maximum library capacities shown in
the figure are calculated assuming LTO9 technology.

Assuming the use of TS1170 technology instead would result in much higher maximum native
capacity, e.g. for the TS4500 library the maximum capacity would be 877 PB instead of 417 PB
with LTO9 technology. (Note that the max number of LTO slots in the TS4500 library is larger
than the max number of enterprise slots due to the slightly smaller size of LTO cartridges.)

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 37. LTO-9 FH Tape Drive and Cartridge

Although somewhat less form factor driven than disk, tape drive offerings have now settled into
the Full Height and Half Height form factors. The very low-end of the tape market that was served
by the 4mm and 8mm tape formats shrunk drastically due to competition from small form-factor
removable disks and the advent of flash memory alternatives, in the late ‘90’s and early 2000’s.
No new low end tape drives have been introduced on the market for many years.

The entry level of the mid-range tape market is served the LTO HH drive which can be installed
in a server, connected to a server via a ‘stand-alone box’ or installed in an ‘auto-changer’ such as
the TS2900 shown in Figure 38 (a). In the middle markets and enterprise sectors, automated tape
libraries have become commonplace and range in capacity from 10s to 1,000s of tapes with
maximum native capacity of a single library now exceeding an exabyte (1 EB = 1018 bytes).

The wide-spread adoption of tape by hyperscale cloud companies has motivated the tape industry
to design solutions to specifically address the unique requirements of this market segment such as
modular designs for ease and speed of deployment as well as optimal reliability in erasure coded
environments. The Diamondback library recently introduced by IBM is one example of this [3].
The Diamondback library can be shipped with media and drives pre-installed and can be installed
in the datacenter in less than 30 minutes. The library is designed for easy self-service, with most
major components replaceable in two minutes or less. It fits in the same floor space as a standard
open compute project (OCP) rack and provides up to 69.5PB compressed capacity using LTO9
technology in a single 8 ft2 (0.7 m2) footprint.

Another recent trend in tape storage is the enablement of object storage to tape. For example,
Spectra Logic offers a tape-based object store with an S3 interface that uses their BlackPearl
gateway. Another example is ActiveScale Cold Storage from quantum that provides an S3
interface to both HDD and Tape based storage with a common name space. Point Archival
Gateway is an example of a software solution that enables a tape-based object store with a
standardized S3 interface and supports tape libraries from multiple vendors as a backend.

This solution also enables erasure coding of objects across multiple tape drives (RAIT) or across
multiple tape libraries (RAIL), an architecture adopted by many hyperscale cloud tape users [4].
Fujifilm also recently announced a software product called Fujifilm Object Archive that provides
an S3 compatible interface to enable object storage with a tape library back end [5]. Objects are
written to tape using an open format developed by Fujifilm called OTFormat [6]. Even more

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

recently, IBM announced Diamondback S3, an S3 based object storage solution that uses the IBM
Diamondback library. There are also open-source initiatives to enable object storage using a tape
back-end such as Open Stack Swift HLM (high latency media), that supports both tape and optical
disc backends [7].

(a) IBM Tape Libraries

(b) Spectra Logic Tape Libraries

Figure 38. (a) IBM Tape Libraries and (b) Spectra Logic Tape Libraries

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

ROADMAP OF QUANTIFIED KEY ATTRIBUTE NEEDS

In 2021, the LTO tape consortium announced the latest generation of LTO magnetic tape, LTO-9.
This tape format supports 18 TB of native storage capacity and advertises a 45 TB cartridge
capacity with 2.5:1 compression. The maximum native drive data rate is 400 MB/s and up to 1000
MB/s compressed, assuming 2.5:1 compression. Like LTO-8, the LTO-9 cartridge format uses
Barium Ferrite (BaFe) particle-based media and tunneling magneto-resistive read sensors. LTO-9
drives can read and write LTO-8 cartridges and support widely used encryption standards.

LTO-9 tape drives operate at an areal density of about 12 Gb/in2 (giga-bits per square inch). On
Aug 23, 2023, IBM announced the TS1170 Enterprise tape drive with 50 TB native capacity, 400
MB/s native transfer rate and areal density of about 26 Gb/in2 using Strontium Ferrite (SrFe) based
media. In contrast to LTO9 and previous enterprise drives, the TS1170 does not provide any
backwards compatibility.

The tape industry began transitioning from the previously used metal particle (MP) technology to
hexagonal platelet shaped BaFe particles around the time frame of LTO-6, with LTO-6 supporting
both MP and BaFe media technologies and LTO-7 media based exclusively on BaFe. The scaling
potential of MP technology was limited by the need for a thin glass coating to prevent oxidation
of the particles. This limited the minimum particle size of MP technology to about 3000 nm 3,
beyond which it is difficult to maintain particle coercivity.

In contrast to MP, both BaFe and SrFe are already oxides and hence do not require a coating and
can therefore be scaled to smaller particle sizes. Both LTO-9 and the IBM TS1160 media use BaFe
particles with a partial perpendicular orientation in the magnetic recording layer of the tape. The
TS1170 media uses SrFe particles and also has a partial perpendicular orientation in the magnetic
recording layer. SrFe is from the same family of hexagonal ferrous oxides as BaFe, but has a higher
saturation magnetization and higher coercivity and hence can be scaled to smaller particle sizes.
LTO tape uses polyester-based substrates whereas the both the TS1160 and TS1170 media use a
more stable but also more expensive aramid-based substrate. The coercivity of both BaFe and SrFe
particles can be tuned over a wide range using substitution elements, similar to doping a
semiconductor.

Hence there is potential to continue reducing the size of BaFe and SrFe particles to enable higher
areal densities while increasing the particle coercivity to maintain thermal stability of the recorded
data. Increasing the coercivity of future tape media will necessitate the use of tape write heads that
produce stronger magnetic fields. This is an area in which tape developers can take advantage of
materials and technologies developed for HDD that currently use media with much smaller
magnetic grains and higher coercivity than tape media.

Table 4 summarizes a set of major tape industry metrics and their expected evolution over the next
15 years. Figure 39 presents the latest LTO Consortium roadmap which describes the five
currently planned future LTO generations [8]. Figure 40 shows the areal density scaling
projections of the 2019 INSIC roadmap for magnetic tape along with historical scaling data for
HDD and tape as well as areal density demonstrations [9]. The LTO roadmap does not provide a
timeline, however, recent LTO generations have been released roughly every 2.5 years. Table 4
projects an areal density of 28.3% CAGR (compound annual growth rate) out to 2031 that will

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

enable a 32% CAGR in capacity scaling. Note that a 32% CAGR corresponds to a doubling in
capacity every 2.5 years, corresponding to the recent trend in time between LTO generations.
Beyond 2031, tape areal density scaling is projected to continue scaling with a somewhat smaller
~27% CAGR and capacity with a 30% CAGR.

Table 4. Magnetic Mass Data Storage Technology Roadmap – Tape

Unit 2021 2023 2025 2027 2029 2031 2033 2035 2037

Native Capacity TB 20 35 61 106 184 321 543 917 1550

Max Data Rate (native) MB/s 400 529 700 925 1224 1618 2140 2830 3743
2
Areal Density Gb/in 12.0 19.7 32.3 53.3 88.0 145.4 233.3 374.8 602.6
Tape Speed (for data) m/s 5.7 6.9 4.1 4.9 5.9 7.1 4.3 5.1 6.1
3
Volumetric Density TB/in 1.5 2.7 4.6 8.1 14.1 24.5 41.4 69.9 118.2
Form Factor* n.a. HH, FH HH, FH HH*, FH HH*, FH HH*, FH HH*, FH HH*, FH HH*, FH HH*, FH

Key Requirements

Read Head type TMR TMR TMR TMR TMR TMR TMR TMR TMR
cpp-GMR cpp-GMR cpp-GMR
Write Head type High Bs High Bs High Bs High Bs High Bs High Bs High Bs High Bs High Bs
Shielded Shielded Shielded Shielded Shielded Shielded
Monopole Monopole Monopole Monopole
Number of channels n.a. 32 32 64 64 64 64 128 128 128
Magnetic Film n.a. BaFe BaFe BaFe BaFe SrFe SrFe SrFe SrFe SrFe
SrFe SrFe SrFe ε-FeO ε-FeO ε-FeO ε-FeO ε-FeO
ε-FeO ε-FeO Sputtered Sputtered Sputtered Sputtered Sputtered
Recording Tech. Perp. Perp. Enhanced Enhanced Enhanced P+SUL P+SUL P+SUL P+SUL
Perp. Perp. Perp. EAMR EAMR
Tape Thickness m 5.0 4.8 4.6 4.4 4.25 4.1 3.9 3.8 3.6

Substrate Material type PEN PEN PEN PEN PEN Aramid Aramid Aramid Aramid
Aramid Aramid Aramid Aramid Aramid Adv. Sub. Adv. Sub. Adv. Sub. Adv. Sub.
Adv. Sub. Adv. Sub. Adv. Sub. Adv. Sub. Adv. Sub.
Notes: Volumetric Density: Capacity divided by volume of an LTO cartridge (4 x 4.1 x 0.8 = 13.12 in 3)
Form Factor: HH = half height = 5.25” half-height internal form factor
FH = half height = 5.25” full-height internal form factor
*Note, due to power density and space constraints HH drives will likely remain in a 32-channel
format with lower data rate when FH drives transition to a 64-channel format.
Read Head: TMR = tunneling magneto-resistive
cpp-GMR = current perpendicular to the plane giant magneto-resistive
Magnetic Film: BaFe = barium ferrite
SrFe = strontium ferrite
ε-FeO = Epsilon iron oxide
Recording Technology: Perp. = perpendicularly oriented media,
Enhanced Perp. = perpendicular media with improved orientation
P+SUL = perpendicular media with a soft underlayer
EAMR = energy assisted magnetic recording
Substrate Material: PEN = polyethylene napthalate,
Aramid = Aromatic Polyamid
Adv. Sub. = Advanced Substrate

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

It is interesting to note that the areal density projected in the last entry of Table 4 for the year 2037
is still about 50% lower than the areal density of current HDD. This provides a level of confidence
that the roadmap projections are attainable. In addition, several research papers demonstrating the
future scaling potential of tape technology have been published in the last few years that provide
further evidence that the projections for the next decade are attainable. For example, in 2017, IBM
in collaboration with Sony, reported a single channel tape recording demonstration of 201 Gb/in 2
using a sputtered tape based on a CoPtCr-SiO2 perpendicularly oriented recording layer [10].

This areal density corresponds to a potential cartridge capacity of 330 TB assuming similar
formatting overheads as an IBM TS1155 drive and considering the increased tape length enabled
by the reduced thickness of the demo media. In 2020, IBM in collaboration with Fujifilm, reported
a single channel tape demonstration of 317 Gb/in2 using a perpendicularly oriented SrFe particle-
based recording layer [11].

This areal density corresponds to a potential cartridge capacity of 580TB for the 4.3 μm thickness
of the demo media and assuming similar formatting overheads as a TS1160 drive. Even more
recently, in 2022, Western Digital studied the recording performance of a sputtered magnetic tape
developed by Sony Media Solutions Corporation that used a sputtered CoPtCr-SiO2 granular
recording layer with a thin CoPtCrB capping layer and reported an areal density of 400 Gb/in2 [12].

Figure 39. LTO Consortium Roadmap [8]

Source: The LTO Program. Hewlett Packard Enterprise, International Business Machines Corp., and Quantum Corporation

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 40. INSIC Tape Areal Density Roadmap and Areal Density of Tape and HDD Demonstrations
and Products [9].
Source: INSIC

CRITICAL ISSUES
Magnetic tape storage will continue to benefit from adapting the head and media technologies
developed by the HDD industry and will have to continue developing advanced servo and signal
processing solutions for the unique challenges that arise in tape systems. In addition, systems level
considerations (performance, error rate, archiving, reliability) will continue to have a major
influence on the design and implementation of tape storage solutions. The rapid growth in adoption
of tape technology by hyperscale cloud companies who buy large quantities of tape but have a
unique set of requirements compared to traditional users of tape will likely influence future product
roadmaps and design tradeoffs between rates of capacity and data rate scaling.

TECHNOLOGY NEEDS: RESEARCH, DEVELOPMENT, IMPLEMENTATION

A critical parameter in continuing to increase areal density in tape drives is the media signal to
noise ratio (SNR). In linear tape recording, the media is the dominant source of noise. The noise
follows particle-counting statistics and is therefore sensitive to the reader width. Modulation noise
and defects also contribute to the noise environment and are significant for sampled amplitude
detection channels such as PRML.

The most effective way to decrease particle noise is to reduce the particle size. However, as the
particles get smaller the capability to maintain coercivity and remanence and the capability to

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

disperse the particles uniformly becomes more difficult. This will continue to be crucial to
advances in tape recording and will consume significant development resources.

Another critical area related to the media is the need to continue to reduce the head-media spacing
to support higher linear densities. Tape is a contact recording technology in which the flexible tape
is run in contact with the tape head and the spacing between the read/write transducers and the
magnetic particles is determined primarily by the surface roughness of the media. Making the
media smoother reduces spacing but at the same time tends to increase tape-head friction which
can degrade recording and servo performance and increase tape and head wear.

To control friction, non-magnetic spacer particles are introduced to reduce the contact area
between the tape and head and hence reduce friction. These particles are also somewhat abrasive
and help to clean debris that might otherwise build up on the tape bearing surface of the head. To
protect the sensitive TMR read transducers, the read elements are coated with a thin wear resistant
coating that also contributes to the spacing.

To achieve the areal densities projected in the later phases of the roadmap will require the
development of ultra-thin wear coatings and very smooth media such that head-tape spacing can
approach that used in current HDD products. In addition, the development of low friction head
technologies that optimize the geometry and topography of the tape bearing surface will be
required to enable the use of such very smooth media. An alternative strategy to address the
spacing challenge is to move away from contract recording and adapt the air bearing technology
developed for HDD to tape recording.

Increasing the track density in linear tape recording will require improvements to track following
servo systems. This includes the process that originally records the servo information on the tape
and the servo system’s ability to follow the tracks on a flexible media. The potential for the
required improvements in tape track following performance have already been shown in the
previously mentioned IBM tape areal density demonstrations that included a track following
component.

In the most recent demonstration, IBM in collaboration with Fujifilm demonstrated a track
following accuracy characterized by the standard deviation of the track following error of 3.2 nm
or less over a tape speed range of 1-4m/s. Here the main challenge will be to implement and adapt
the lab technologies used to achieve this performance in commercial tape drives and media
manufactured at scale.

The requirement to increase cartridge capacity will lead to thinner tape, which increases the
difficulty of handling and guiding the tape. The introduction of flangeless tape paths significantly
reduced the potential for tape edge damage; however, thinner media will necessitate the need for
improvements in tape path mechanics and tension control during tape transport.

The dimensional stability of the tape is a large part of the off-track budget for a multi-channel tape
system. The recent introduction of active tape dimensional stability compensation has relaxed the
need for continual improvement in the dimensional stability of tape substrate materials, however
continual improvements in the accuracy of the compensation mechanism will be required to enable

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

continued track density scaling. While not trivial, active TDS compensation shares many of the
same challenges, uses the same position measurement signals and hence benefits from advances
in track following technology. The nanometer scale accuracy that has been demonstrated for tape
track following control provides some confidence that similar levels of accuracy are attainable for
TDS compensation.

Beyond automating backup and archiving applications, tape libraries have enabled the use of tape
in hierarchical storage management (HSM) applications for improving performance with the
migration and recall of data sets in a variety of storage environments. Data storage management
hardware and software, including virtual tape, has enabled increased utilization of the tape
cartridge while improving the performance and retrieval times for the tape environment by
combining faster access storage such as HDDs or flash memory to the tape storage system.

Archival tape storage has evolved in a similar direction, with the development of automated high-
performance tape storage systems having native capacities larger than one exabyte (EB) and
achieving data center floor space densities of up to 4.77 PB/ft2 (TS4500 library with TS1170 tape
technology). Magnetic tape continues to offer the highest probability of success in achieving the
maximum total volumetric density with acceptable data rate combined with the lowest price per
terabyte of any storage technology. The improvement in tape usage, drive and media reliability
has made tape acceptable in a wide variety of high-capacity, moderate to low activity applications.
Fixed content, archive, compliance and disaster recovery applications are the primary uses of
magnetic tape today as the traditional backup and restore market has largely moved to magnetic
disk-based architectures.

Linear tape cartridges have sustained about 30-40% annual growth rate in storage capacity over
the last decade through the combination of breakthroughs in four areas. These are the incorporation
of magneto-resistive, giant magneto-resistive and tunneling magneto-resistive heads, advances in
track following servo technology, advanced in ECC and data channels as well as advances in media
technology. Several years ago, several tape cartridges were needed to back up a single disk drive;
today a single tape cartridge can often backup multiple disk drives.

Today tape remains the greenest, i.e. the most energy efficient and lowest CO2e, of all mass storage
technologies. These attributes are expected to increase the economic appeal of tape systems in the
coming years.

GAPS AND SHOWSTOPPERS

Continuing to achieve the required head tracking precision and the continued advancement of
media are two key challenges for tape’s continued progress. Progress in the ease of use and
integrations of tape into hierarchical storage systems combined with continued areal density
growth during the recent slow-down in areal density scaling of HDD has created an opportunity
for tape to increase the capacity and cost advantages of tape storage over HDD.

New tape architectures are addressing indexing, tags and naming conventions and enabling a move
toward an “object-oriented” approach to keep the vast storage reservoirs tape provides usable.
These concepts benefit from more intelligent drive and cartridge systems that can rapidly locate
specific objects or data, enabled by technologies such as the LTFS file system. All the tape drive

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

vendors now have file systems access to tape, effectively creating Network Attached Storage
(NAS) tape. In addition, many vendors are moving to object storage support on tape. Object storage
with magnetic tape in cloud-based storage environments is a future growth opportunity for tape
archiving.

RECOMMENDATIONS ON PRIORITIES AND ALTERNATIVE TECHNOLOGIES

Tape storage will continue to play a key role as the foundation of the storage hierarchy “pyramid”,
providing an archival non-volatile solution for vital data records of businesses and governments,
as well as (indirectly) for consumer data stored in “the cloud”. As such, the underlying magnetic
recording, as well as drive and library systems technologies, needs to be continuously tracked to
understand and follow improvements in performance and reliability. Of particular interest will be
how other storage technologies such as hard disk drives and solid-state drives may be used to
augment and enhance tape systems performance and efficiency. Moreover, adding a file system
and object storage capability to tape storage systems has had a dramatic impact on the design of
tape library systems and their use. The large-scale adoption of tape technology by hyperscale cloud
companies is also likely to influence future development directions for tape drives and libraries.

REFERENCES
1
A. Acilla, “Quantifying the Economic Benefits of LTO-8 Technology” Accessed: Dec. 3, 2022. [Online].
Available: https://www.lto.org/wp-content/uploads/2018/08/ESG-Economic-Validation-Summary.pdf
2
IDC Worldwide Branded Tape Share Report Four Quarter Roll-up through 4Q 2021
3
IBM Diamondback Tape Library, Accessed: Dec. 23, 2022. [Online]. Available:
https://www.ibm.com/products/diamondback-tape-library
4
“Point Archival Gateway”, Accessed: Dec. 3, 2022. [Online]. Available:
https://www.point.de/en/products/point-archival-gateway/
5
“Object Archive”, Accessed: Dec. 3, 2022. [Online]. Available: https://www.fujifilm.com/us/en/business/data-
storage/data-management/data-archive/object”
6
“OTFormat Specification”, Accessed: Dec. 3, 2022. [Online]. Available: https://activearchive.com/wp-
content/uploads/2020/09/OTFormat_Specification_ver1.0.0-1.pdf
7
Github repository for SwiftHLM (Swift Hight-Latency Media) middleware, Accessed: Dec 3, 2022. [Online].
Available: https://github.com/ibm-research/swifthlm
8
Linear Tape Open Roadmap, Accessed: Dec. 3, 2022. [Online]. Available: https://www.lto.org/roadmap/
9
Information Storage Industry Consortium (INSIC) 2019 International Magnetic Tape Storage Roadmap,
Accessed: Dec. 3, 2022. [Online]. Available: http://www.insic.org/wp-content/uploads/2019/07/INSIC-
Technology-Roadmap-2019.pdf
10
S. Furrer et al., “201 Gb/in2 recording areal density on sputtered magnetic tape,” IEEE Trans. Magn., vol. 54,
no. 2, Feb. 2018, Art. no. 3100308. DOI: 10.1109/TMAG.2017.2727822
11
S. Furrer et al., “317 Gb/in2 Recording Areal Density on Strontium Ferrite Tape” IEEE Transactions on
Magnetics, JULY 2021, 57, 7, pp. 1-11, DOI: 10.1109/TMAG.2021.3076868
12
P.-O. Jubert, Y. Obukhov, C. Papusoi, and P. Dorsey, “Evaluation of Sputtered Tape Media with Hard Disk
Drive Components”, IEEE Transactions on Magnetics, Vol. 58, No. 4, April 2022 , DOI:
10.1109/TMRC53175.2021.9605133

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Optical Archival Data Storage

SITUATION ANALYSIS
Optical discs are a nearly ubiquitous form of digital data storage as over 256 billioni units [1] have
been sold over the last 40+ years (compared to 10 billion HDD and 352 million LTO units). CDs,
DVDs and BR discs comprise the media collections of consumers all over the world including
CDs that are still playing pristine, high-fidelity music after four decades. Perfect fidelity DVDs
have also demonstrated decades of longevity.

CDs and DVDs have been shown to survive the most extreme environments of up to 80°C when
stored in automobiles. This outstanding longevity and robustness to environmental conditions
make optical storage an important approach for the burgeoning need for a carbon-friendly and
energy efficient enterprise active archival data storage medium.

The previous applications of optical discs have included replicated discs for media distribution and
write-once read many (WORM) and rewritable technologies for data storage, mostly for consumer
applications. This report will focus on the recent developments in WORM optical media and their
forward-looking roadmaps toward applications in enterprise archival storage. Following a brief
summary of the historical consumer-oriented technology and applications, the value proposition
of optical storage for enterprise archival data storage will be described.

The incumbent technology is characterized by optically written data marks on thin layers in single
and multiple photosensitive layers deposited on plastic substrates. Discs are spun at high speed
with focus and tracking servos securing rapid access to data locations in three dimensions.

Multilayer disc media will be reviewed, while new structures in the market and in development
will be reviewed in this report, including a description of the media, drives and libraries comprising
the current and future optical media technologies.

CONSUMER MARKET
Research into optical data storage and its practical applications has been ongoing for many
decades, with analog image microfiche widely considered as the first optical storage medium,
introduced as early as the 1800s [2]. The first true widely-adopted digital data storage system was
the replicated Compact Disc, introduced in 1982, adapted from audio (CD-DA) to data storage
(the CD-ROM format) with the 1985 Yellow Book, and re-adapted as the first mass-market optical
storage medium with CD-R and CD-RW in 1988. Subsequently, DVD and Blu-Ray discs have
been developed in the 1990s and 2000s respectively to accommodate full-length movie distribution
in ever-increasing display resolution [3].

CD/DVD/BLU-RAY
The original concepts of CD/DVD drives continue to be the basis for the incumbent multilayer
enterprise data storage media, with innovations built on the basics: dynamic focus and tracking
servos and PRML and related data channels, among others. Indeed, the disc form factor is common

i
Estimate based on Global Consumer CD unit sales and Global Consumer DVD and BD $ sales

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

to all the multilayer technologies described below. Digital data is encoded with marks of various
lengths with accepted error correction codes. The evolution of the layered disc toward higher
capacity has relied on shortening the laser wavelength and increasing the numerical aperture of the
objective lens to enable shrinking the size of the diffraction limited spot and shortening the track
pitch, as depicted in Figure 41. Other innovations described below also increased the data
density/disc capacity. Capacity increases also involved the addition of layers.

Since the focus of CDs and DVDs was mass media distribution, few improvements were made
throughout their lifecycle, and the development effort has been mostly focused on standardization
of manufacturing and cost reduction. The transition from CD to DVD was mainly driven by the
storage of feature films versus music collections. The transition to Blu-ray was driven by increases
in video display resolution.

Figure 41. Decrease in laser wavelength and spot size with increasing numerical aperture (NA)
lenses. [4].
Reproduced with permission: American Chemical Society

Blue-Laser Disc is defined by Blu-ray Disc (BD) format and all other types of optical data storage
technology (for example, UDO) that use blue lasers for writing/reading. Thus, Blu-ray became the
natural successor to DVDs and HD DVDs with its superior capacities, (25 GB per layer vs 4.7 GB
per layer) which were essential to distribute 4K and ultra-high-definition videos.

The introduction of Blu-Ray coincided with the popularization of streaming media over the
internet. With competition from both the lower cost of DVDs and the improved convenience of
streaming and digital downloads from vendors, Blu-ray never experienced the same level of
success as DVDs and CDs. Blu-ray sales only barely surpassed DVDs sales recently, which was a
shadow of its former volume two decades ago. The trend away from consumer applications of
optical media is depicted in Figure 42[5].

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Personal computer applications of optical storage have also been a traditional market for optical
media. In this case, the on-board drives have been used for software distribution on read-only discs
as well as local data storage on writable and re-writable discs.

Traditionally, read-only CDs and DVDs (CD-DA, CD-ROM, VCD, DVD-ROM, BD-ROM, etc.),
are used for large software distributions. However, with the advent of downloadable software and
increased broadband availability, optical discs became a less popular or desirable means of
distributing software [6], [7]. More recently, major gaming consoles are releasing devices without
optical drives despite the piracy concerns associated with the industry [8].

Figure 42. Music delivery. DVD sales exhibit a similar trend to CD sales with a peak in the year
2004 [9], [10].

For a time, write-once and rewritable discs (CD-R, CD-RW, DVD+/-R, DVD+/-RW, DVD-RAM,
BD-R, BD-RE, etc.) were commonly used for general file operations using simulated file systems
such as UDF [11]. However, this market declined quickly with the introduction and widespread
use of NAND-based SSDs. The performance and capacity of SSDs make them the obvious choice
in these applications.

DISCONTINUED/MARGINAL CONSUMER MEDIA

Very few major developments in the optical storage space pertaining to consumer applications
took place over the past 5 years. Please refer to the previous edition of the iNEMI roadmap [12]
for more details on technologies such as Magneto-Optical, VMD, Millennium Discs, and others.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

ENTERPRISE USE CASE VALUE PROPOSITION

Just as the demand for archival data storage is rapidly expanding due to e.g. data science demands
and the internet of everything, both business and technical issues are straining the incumbent
magnetic media. These stressors include the breakdown of Kryder’s law [13], the necessity for a
sustainable approach to data storage, and the oligopoly of both storage producers and consumers.

Figure 43. Storage media comparison. [14]

Reproduced with Permission: Panasonic Holdings Corporation

Optical data storage has the potential to meet these challenges. Several elements of the value
proposition of optical data storage are shown in Figure 43[15]. These advantages arise from the
robust nature of marks produced by WORM photothermal processes in the photosensitive
materials comprising the active layers. As the photothermal conversion occurs at hundreds of
degrees, the resulting WORM marks are extremely robust, leading to a long lifetime as the disc
materials also have century-scale lives. Figure 44 depicts the lifetime/temperature trade-off for AD
technology indicating that even at the unlikely occurrence of continuous storage at 35 C, decades
of life are expected [16], [17]. The environmental robustness and longevity have been verified over
the decades since the introduction of optical discs as mentioned in the introductory paragraph.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

These generic attributes of optical media also minimize power consumption and carbon footprint,
which is increasingly important for enterprise archives facing increasingly larger cold data storage
sizes. Unlike magnetic media, optical media can generally be stored in relatively uncontrolled
ambient environments, leading to significant energy/cost savings as indicated in Figure 44 and
Table 5. As a result of this and their longevity, optical media generally require zero energy to
store, an important factor in their energy consumption and carbon footprint.

Figure 44. Accelerated lifetime test results for Sony/Panasonic Archival Disc. [18], [19]
Reproduced with permission: Sony Group Corporation and Panasonic Holdings Corporation

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Table 5. Common Enterprise Archival Storage Media

Unit Media Transfer BER Archiving Power TCO (Total Cost
Capacity (Cost/TB) [20] Speed with Error Consumption of Ownership
(MB/sec) Corr. (/time/Capacity)ii ($MM/100GB/20
years)iii
HDDiv 22 TB >$10 291 1E-15 ~440 7.6
2.2
TB/disk
Tapev 18 TB >$5 400vi 1E-19 ~41 2.9
(Native)
BDXLvii 100 GB <$40 18-36 1.00E-18 ~14 0.7
to 200
GB
Archival 0.5 TB $30-45 30-45viii 1E-23 ~14 <0.7
Disc

The long lifetime is especially important for data archiving for decades as the media does not need
to be remastered. Typical remastering intervals for HDD and magnetic tape are 3-10 years, so that
archiving data for 50 years can result in an order of magnitude of cost savings accounting for both
media and remastering infrastructure.

The robustness of optical media has important cybersecurity implications as the WORM data is
impervious to tampering. Because the discs are stored separately from the drive, the system
possesses an “air gap” to provide passive media protection from cybersecurity breaches. In
addition, optical media are impervious to electromagnetic pulse (EMP) and immersion in water.

Current commercial optical data storage technologies are based on concepts developed almost 50
years ago for consumer applications. The pivot of optical storage to data center archiving presents
opportunities for new technical approaches at the system level, allowing novel concepts at the
device/media level as well. In addition, advances in optical components driven by the imaging and
display industries introduce new components for such systems. These advances will be described
below.

TECHNICAL REQUIREMENTS (BUSINESS/TECHNICAL ISSUES)

Despite its advantages, optical archival storage has not found widespread use as neither its capacity
nor its cost is competitive with HDD and tape. At 1,500 PB per year, the optical archival market
is less than half of 1% of the total storage market.

ii
Optical Technology | Optical Data Archiver freeze-ray series | Panasonic Global
iii
Low Power and Cost Efficient Data – HIE Electronics
iv
WD Gold Enterprise Class SATA Hard Drive Up To 22TB | Western Digital
v
Fuji LTO Ultrium 9 Tape (16659047). Fujifilm LTO9 Ultrium Tape Data Cartridges with Barium Ferrite (BaFe)
(tapeandmedia.com)
vi
Transfer speed for a tape cartridge, which leverages multi-track parallel transferring. LTO-9 models can write on
32 data tracks at a time.
vii
Disk Prices (US)
viii
Transfer speed per individual optical discs. Higher transfer speed can be achieved through multi-unit drives and
various library systems. See Archival Disc Libraries Section for detail.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Optical storage technologies have evolved from CD to DVD to Blu-ray as capacity requirements
have increased for music, video, gaming, and personal storage. The basic disc and drive technology
have not undergone fundamental change. On the disc side, the basic disc format has remained
constant with track density reflecting the increasing capacity permitted using higher numerical
aperture lenses and shorter wavelength lasers.

The development of wide bandgap semiconductor lasers such as those using GaN, the subject of
the 2014 Nobel Prize in physics, was a major innovation leading to 405 nm wavelength lasers for
Blu-ray and, now, Archival Disc (AD) technology. Sony and Panasonic introduced a two-sided
disc along with improvements in material and drive technologies to achieve the significant capacity
advances in Archival Disc technology. Drive improvements include progress in PRML, other data
channel implementations, and error correction schemes.

Generations of optical discs have shared an additive manufacturing process consisting of an

injection molded polycarbonate substrate with the tracking pattern embossed on the upper surface.
For CD and DVD technology, the relatively low numerical aperture of the objective lens provided
a large working distance so that a polycarbonate slab can be used as the cover of the written layers.
The transition to the blue laser and 0.85 numerical aperture required a shorter working distance,
well below the 1mm scale.

This required the development of both a deposited thin cover layer and a hard protective coat.
Various materials have been used as active media over the years, but current blue laser re-writable
technologies are based on photothermal inorganic phase change composite materials that change
from reflective crystalline to non-reflective amorphous states upon writing. Typically, phase
change composites contain Te, Ge, Sb along with other elements.

On the other hand, for WORM technology like recordable Blu-ray discs, various inorganic
materials such as phase-change composite, metal-oxide and metal-alloy have been used. These
inorganic materials are UV and temperature stable and have long lifetimes, making them suitable
for archival applications. Alternatively, photosensitive organic dyes have been used for recordable
Blu-ray discs using Verbatim’s spin-coating process instead of the traditional sputtering process
for inorganic materials.

The capacity of Blu-ray discs can be potentially increased by adding more recording layers [21],
but an additive, sequential layer deposition process even at 98% yield for a single layer would
result in a prohibitively low manufacturing yield for more than 3 or 4 layers. Table 6 is a summary
of some current and novel optical products and technologies.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Table 6. Summary of select current and novel optical products/technologies

Max Media Read/Write BER
Product Development Service Capacity Cost per Speed with Error
Name State Recording Medium Segments GB/disc TB (MB/sec)ix Corr.
Photosensitive organic 1.00E-9
CD/DVD Commercial dyes Consumer 0.7-30 $15-$30 7.2-66/7.2-66 1.00E-15x
Phase change inorganic ~1.00E-18,
composites; photosensitive Consumer Varies by
Blu-ray Commercial organic-dyes Enterprise 50 $15+ 36-72/36-72 drivexi
~1.00E-18,
Phase change inorganic Consumer Varies by
BDXL Commercial composites Enterprise 200 <$40 18-36/9-36 drive
Archival ~1.00E-
Discxii Commercial Inorganic oxides Enterprise 300-500 <$33 90 23
Photosensitive dyes
dispersed in polymer
Folio matrix – reflective or 500+
Disc Prototyping fluorescent media Enterprise (projected) <$5 TBD TBD
polyester photographic
Piqlxiii Commercial film Enterprise N/A N/A 40 TBD
metallic alloy sputtered on
polyester photographic 1200+
DOTS Prototyping film Enterprise (projected) TBD 1,000 TBD

The photosensitive material possesses a strong optical absorption feature at the 405 nm wavelength
so that the absorbed optical energy of the writing pulse is quickly converted to heat which increases
the temperature beyond a phase transition or decomposition event. Both processes, then, are
characterized by an absorbed-heat threshold converting the active material to produce the mark.
The threshold nature provides that the much-lower reading power level of the incident laser does
not significantly affect the written marks, thus ensuring millions of reading cycles. This is also an
important factor in the photostability of the written data to ambient light.

The production of data marks is a complex process, involving not only the photothermal event but
also the transport of heat over the relevant time and length scales. The mark is written with a
particular writing strategy, which is a sequence of laser pulses near the nanosecond time scales
carefully crafted to minimize thermal transport and create the smallest marks, at best at the ~10-
100 nm length scale in blue laser media, well below the diffraction limit of light at 405 nm
wavelength. The writing strategy depends not only on the materials, but also on the dimensions of
the layers as they affect heat confinement and transport.

INCUMBENT MEDIA: MULTILAYER DISC

BDXL
The Blu-ray Disc Association released specifications for two multi-layer variants of Blu-ray discs
in 2010, dubbed Blu-ray XL (BDXL) [22]. With this release, BDXL increased Blu-ray capacity to
100 GB (rewritable) or 128 GB (WORM). The fourfold increase in capacity is achieved by using
three or four recordable or rewritable BD layers without increasing their areal density. This opened

ix
Pioneer Just Made A New Optical Disc Drive For PCs And Yes It's 2022 | HotHardware
x
DVD Benchmark (hometheaterhifi.com)
xi
Error-correction codes for optical disk storage [5643-58] (psu.edu)
xii
Folio Photonics Wants To Kill HDDs With Film (forbes.com)
xiii
https://www.ejournals.eu/pliki/art/20806/

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

the door for Blu-ray disc enterprise archival applications, which have gained traction in some
sectors.

In 2014, a double-sided version of Blu-ray standard was released (BD-DSD), which allows 200GB
of storage between 6 layers (3 on each side) [23]. However, BD-DSD requires specialized optical
drives to process and is not widely available. Since there is currently no roadmap for BDXL
development, this solution will likely be capped at the current capability.

In March 2022, Pioneer released a new optical disc drive that delivers 8x/6x speed recording on
triple/quadruple-layer BDXL discs, a substantial improvement over the 4x/2x speed drive released
in 2017 [24]. It is unclear whether there will be an ongoing roadmap moving forward. Figure 45
shows the layer structure for 3 and 4 layer BDXL discs.

Figure 45. BDXL layer structure. [25]

Reproduced with permission: Blu-ray Disc Association

SONY + PANASONIC ARCHIVAL DISCS

Sony and Panasonic have successfully developed a next-generation optical disc for enterprise
storage with an initial capacity of 300GB (for 2016 shipment). The Archival Disc (AD) has the
same dimensions as current Blu-ray discs and will also be readable for at least 50 years. The disc
has three layers per side.

Figure 46. Land-and-groove recording. Blu-ray Disc on the left vs. AD disc on the right. [26]
Reproduced with permission: Sony Group Corporation and Panasonic Holdings Corporation

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

A major breakthrough of the Archival Discs was the utilization of land-and-groove recording
technology, nearly doubling the areal density per layer compared to conventional Blu-ray discs
(Figure 46). This also enables the Archival Discs to use existing optical units with the same 405nm
laser at 0.85 NA. Sony and Panasonic indicate that the crosstalk noise generated between adjacent
tracks is cancelled out by their newly developed crosstalk-cancelling technology. AD is backward
compatible with existing Blu-ray standards. Panasonic AD drives also support BD discs.

Sony and Panasonic have been steadily making improvements on the AD product. In 2020, Sony
and Panasonic released the second generation of discs with 500GB capacity, which was achieved
primarily by shortening the data bit-length, thus improving linear recording density [27]. In
addition, the use of new oxide-based materials improves recording rate, disc capacity and
durability. This improvement was made possible by AD’s advanced inter-symbol interference
elimination technology, which will be fitted to 2nd generation drives to rectify reduced playback
spot resolution caused by higher recording density. This generation of AD also benefited from a
new data format, which improved linear recording efficiency by 7%, and new channel
modulation/advanced error correction code to reduce data error rates.

Their literature describes the next generation to have an estimated capacity of 1TB and indicates
that this higher capacity will be achieved through signal-processing technologies including multi-
level recording technology as depicted in Figure 47.

Figure 47. 3rd generation AD multiple-level recording/playback performance [28]

Reproduced with permission: Sony Group Corporation and Panasonic Holdings Corporation

Sony and Panasonic’s 3rd generation AD would double the capacity of 2nd generation by encoding
twice as much data with the same format and physical disc structure by effectively encoding 2 bits
of information per each data spot (Figure 48). Over the course of each generation, disc capacity is
raised without changing base optical parameters or media structure, minimizing manufacturing
cost and ensuring the AD system’s backward compatibility.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Sony and Panasonic note that Archival Discs would last 50 to 100 years archival life depending
on various environmental factors. The discs have been demonstrated to be readable after being
immersed in seawater for a period of 5 weeks. Sony also states that since optical discs are not
magnetic, they are unaffected by geomagnetic events and other sources of electromagnetic pulses
(EMP).

Figure 48. Archival disc capacity roadmap. [29]

Reproduced with permission: Sony Group Corporation and Panasonic Holdings Corporation
Sony and Panasonic depict the roadmap for higher-capacity ADs. As the graphic in Figure 48
shows, the third generation will have a 1 TB capacity. Sony and Panasonic have not disclosed
plans for higher capacity beyond 1TB. Consider the data in Table 7. Here the basic disc specs are
given.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Table 7. Main parameters of AD discs. Reproduced with permission: Sony Group Corporation and
Panasonic Holdings Corporation [30]

DATA PRESERVATION
New products on the market and in development are focusing on immutable storage at the
millennium scale. These will address preservation markets and may also be applicable in markets
for more active data storage. Piql AS has products in the market and more are described later as
under development.

PIQL AND PIQLFILMxiv

Backed by the European Union and the Norwegian government, the Norwegian company Piql has
developed a photosensitive film, piqlFilm, that both protect and preserves data for 1000 years [31].
What Piql has done is in principle to convert an established and well proven information carrier,
the 35 mm black and white photographic film, traditionally used for analogue data storage, into a
digital storage and preservation medium that can be used to store any digital data [32]. The
technology preserves data in the form of high-resolution QR codes that is decodable using open-
source software. The piqlFilm has been developed in collaboration with three film manufacturers,
Kodak (US), Harman Technology (UK) and Filmotec (DE). Figure 49 shows the piqIFilm
package

Figure 49. piqlFilm package

xiv
This section authored by Piql AS Managing Director Rune Bjerkestrand

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

The piqlFilm is a black and white, negative silver-halide film on polyester base, see Figure 50. It
has extremely fine grains (20 nm to 40 nm) and high resolving power (>1000 line-pairs per mm).
Further, piqlFilm has truly unique security and longevity properties, ideally suited for offline/“cold
storage” of valuable and/or irreplaceable data and information. Data stored on piqlFilm cannot be
hacked, modified, or deleted, neither can data be destroyed by electromagnetic weapons or nuclear
radiation. The piqlFilm is packed in the piqlBox to protect it physically and over time. The piqlBox
is manufactured using a specially designed and manufactured polymer material that has no
negative impact on the piqlFilm and data over time and that has the same longevity as the piqlFilm,
i.e. 1000 years. The same goes for the label used on the piqlBox.

The piqlFilm is made self-contained and self-explanatory due to the fact that it contains human
readable instructions (i.e., in addition to the digital data, see Figure 51) on how to understand the
storage medium and how to deal with it in the future. Further it contains all file format descriptions
and relevant source code for programs needed to render or view information in the future. Piql has
even developed a virtual machine that makes future data retrieval independent of specific hardware
or operating systems available at that point in time. This makes the solution resilient against the
accelerating developments and obsolescence of specific software and hardware.

Figure 50. Digital and analog data on piqlFilm

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 51. piqlFilm. Reproduced with permission: Piql AS

Since the piqlFilm is migration-free, passive and requires no energy to keep data alive, it is a truly
sustainable storage technology with a close to zero carbon footprint over time.

Data is ingested through Piql’s software platform, piqlConnect, where data upon being organized
by the client, is transferred to a specific machine, the piqlWriter, that encodes the received bit
stream into QR-codes, such as shown in Figure 52.

Figure 52. Sample binary data on piqlFilm

The piqlWriter receives the encoded binary data as a set of files representing the writable data-
frames (images) and writes them on the piqlFilm at 20 data-frames per seconds, or roughly
40MB/s. Imaging is based on a Texas Instruments Digital Micromirror Device (DMD) imaging
sensor, consisting of more than 8,800,000 micro-mirrors capable of writing datapoints of 6μm size.
The piqlWriter uses a monochromatic green LED light which is modulated by the DLP. Each
datapoint can be written in black, white or shades of grey.

The photographic nature of the film enables piqlFilm to store both analog and digital data, which
is essential for ultra-long-term data preservation purposes, where future data retrieval without

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

current technology is an important consideration (Figure 50) [33]. The decoding method is written
in human-readable text that can be read with a magnifying glass, and the digital data stored is
searchable as with any other digital media.

One fundamental principle for Piql when they designed their solution was to ensure that data could
be read back tomorrow, next year or a thousand years into the future, without the need for specific
hardware or software. What in principle is needed to retrieve the data from piqlFilm is:
• a light source (to illuminate the data-frames on the piqlFilm),
• a magnifying glass (to hold over the piqlFilm where illuminated),
• a camera (that can capture the image (i.e. the QR code) as seen through the film and the
magnifying glass),
• and a computer (that can interpret and convert the QR code into the original file format).
All the instructions, the file format descriptions and the software that is needed (i.e. the source
code), is all included on the piqlFilm and can be read by the human eye and understood by a non-
technical person. Even the instructions how to build an automated reading device is included on
the piqlFilm.

To automate and industrialize the readback of the piqlFilm, Piql has developed the piqlReader
(Figure 53).
The piqlReader is a high speed, high-resolution digital data reader. It captures data-frames from
the film and restores digital and visual data. The piqlReader is used for two purposes; The first use
is data verification after the film has been written and subsequently processed. This ensures that
data written on film is verified to be both readable (with low error rates) as well as authentic (i.e.
checksum of the files are verified). The second use is data retrieval upon request. When a client
requires data from the film, the piqlReader is used for accessing the data.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 53. A piqlReader

The piqlReader features precision mechanical film handling components, customized optics, a
diffused 405nm LED light source, a built-in line-scan 12K camera and a high-performance
computer. It runs open-source software (available on GitHub) for capturing images and decoding
them to digital data in real-time. Retrieving digital data from film requires two processes running
in parallel. The first process is capturing the image from the film, and the second is decoding that
captured image file. The piqlReader reads in continuous mode and can quickly access the needed
file, whether at the beginning or at the end of a piqlFilm. Data retrieval speed is 24 MB/s or 12
data-frames per second.

Piql has established a global network of distributors; Piql Official Resellers, Piql Partner and B2B
partners (i.e. larger system integrators/solution providers) that deliver Piql’s services for data
protection, archival and long-term digital preservation. Clients across the world can reach
these services through the SaaS platform, piqlConnect, that is the connection to/from the offline,
off-grid piqlFilm as well as online storage. The Piql Partners (located across the world on all
continents) manage a local piqlVault where the piqlFilm is securely stored.

Piql’s services have been delivered to prestigious clients like the Vatican Library, European Space
Agency, GitHub, various National Archives, National Museums, National Libraries, banks, large
corporates, SME’s, research institutions, various public agencies (legal, infrastructure, utilities,
nuclear, health, defense, public administration and more).

Piql is also the initiator of the Arctic World Archive (AWA), a repository for World Memory
located on the Svalbard archipelago in the Arctic Ocean, between the top of Norway and the
North Pole.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

AWA was established in 2017 [34]. based on a vision to ensure that our digital memory is not
lost or manipulated but shall be available for future generations in its authentic form, - in a world
where few places are safe from natural and man-made disasters, hacking, cyberattacks, wars, and
terror. Data is stored on the piqlFilm that is kept in a secure vault in the dept of the permafrost in
an Arctic Mountain. The lifetime of the data is expected to be beyond 2000 years when stored in
a cold and dry climate like in the Arctic World Archive.

OPTICAL STORAGE SYSTEMS

BDXL LIBRARIES
Amethystum, KDS, and NETZON
Some Chinese archival storage management system providers have created library systems that
integrate optical, magnetic, and NVM storage media in integrated systems to achieve cost-
performance optimization in enterprise storage. This approach allows the volume density to be
comparable to other storage solutions, while streamlining optical-magnetic-digital system
integration. Major library makers include Amethystum, KDS, and NETZON (Figure 54). These
solutions are popular in the Chinese market.

Figure 54. NETZON HDL 10368 holds 10368 discs enclosed in 864 cartridges and 36 parallel
drives. [35]
Reproduced with permission: Suzhou NETZON Information Storage Technology Co., Ltd.

In the realm of pure optical media library development, the library makers take a similar approach
[36]. BDXL library makers pack as many discs as possible, with up to ~12000 Blu-ray XL discs
packed in a cabinet, realizing a 2.5 PB/cabinet deployment (1.25 PB for the standard 19 “42U
cabinet) [37]. Multiple drives allow for high parallel data access speed. Table 8 shows
specifications for the Amethystum ZL BDXL Optical Storage System.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Table 8. Amethystum ZL series BDXL Optical Storage System Specifications [ 38]

Model ZL600 ZL1800 ZL2520 ZL6120 ZL12240
Storage 60TB 180TB 504TB 1224TB 2448TB
Disc number 600 1800 2520 6120 12240
Maximum Drive Number 6 6 12 24 48
Network Interface 1Gb/10Gb Ethernet
Avg. grab time 14s 14s 60s 60s 60s
Max Transmission rate 162MB/s 162MB/s 324MB/s 548MB/s 1296MB/s
Voltage 100-240V AC / 47-63Hz
Size 19-inch 24U 19-inch 37U 19-inch 25U 19-inch 42U 1000x800x20
cabinet cabinet rack cabinet 90mm
Weight (Fully Loaded) 150kg 240kg 180kg 454kg 1028kg
Operating Environment 10°C to 35°C, 20% to 80% Humidity

Consumer Library Designs for Enterprise Use

Some consumer-grade optical archival systems (Optical Jukeboxes) manufacturers such as Zerras
and Kintronics have developed enterprise solutions by upscaling their consumer products [39],
[40]. These lower capacity libraries have short (<10 seconds) access time and the minimal number
of robotic mechanisms may prove desirable for near-line storage applications that rely on the
flexibility and modularity it provides.

ARCHIVAL DISC LIBRARIES

The technology for a modern data archiving system based on optical discs requires two primary
elements: a robust, high-capacity optical disc and system architecture suitable and effective for
enterprise data centers.

The collaboration between Sony and Panasonic on the Archival Disc has established system-level
products, but the high cost of the media has hobbled market penetration. The systems described
below have recently been discontinued according the Sony and Panasonic websites.
Sony Everspan
Everspan aimed for an ultra-high-capacity storage solution, but has been displaced by Sony
PetaSite [41]. It consisted of three types of units: The Base Unit, the Robotic Unit, and up to 14
Expansion Units, based on the ‘triplet’ rack developed by the Open Compute Project (OCP) and
containing up to 43,520 Archival Disks each. When using 300GB ADs, the total capacity of a
single Everspan system is 181 Petabytes and up to four systems can be linked. Everspan has a
relatively high I/O rate – the robotic read-write array features not one, but eight lasers for a total
read speed of up to 18 Gigabytes per second.

Sony PetaSite
In 2020, Sony released the 3rd generation of Optical Disc Archive Cartridge alongside the
corresponding drives [42]. The third-generation disc cartridge holds 11 AD 500GB discs with a
maximum capacity of 5.5TB, compared to the previous generation’s 3.3TB. The third generation
ODS-D380U/F features eight lasers. With two assemblies positioned at the top and two at the
bottom, the drive can read and write both sides of the disc simultaneously. The drives were able

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

to improve the read/write speed to 375MB/187.5MB (3Gbps/1.5Gbps) while maintaining the

backward compatibility of the 1st and 2nd generation cartridges. Figure 55 and Table 9 show the
Sony PetaSite library system and give the library system specifications.

Figure 55. Sony PetaSite library system. [43]

Reproduced with permission: Sony Group Corporation

Table 9. Sony PetaSite library system specifications. [44]

Reproduced with permission: Sony Group Corporation

Sony’s latest Optical Disc Archive, PetaSite Scalable Library, return to the traditional 42U rack
form factor [45]. Sony achieves this scalability by deploying a “Master Unit” and up to 5
“Extension Units”. The ODS-L30M forms the initial building block of the PetaSite modular library
solution. It provides robotics for an entire rack with support for two drives and 30 cartridges. To
increase archive performance and capacity, the ODS-L60E modular extension unit supports up to
4 additional drives and 61 cartridge slots. If additional archive capacity is all that is required, then
the ODS-L100E can be added to provide an additional 101 slots of cartridge capacity. Note the

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

PetaSite Library uses the ODA optical drives, which read/write AD discs without removing them
from their individual cartridges, essentially mitigating the dust issue presented in the Everspan
system. Note this solution is also superior in terms of environmental tolerance, featuring a wide
range of operating conditions that minimize energy cost from air conditioning. With 4 drives
installed in an ODS-L60E library, a total of 750MB/s per rack transfer speed can be achieved.

Panasonic Freeze-ray
Panasonic has a system configuration similar to that of Sony’s, with the key difference being that
discs are handled and stored in magazines, and discs are written/read individually after being
extracted from the magazines mechanically [46]. Concurrent read/write is possible with multiple
drives per unit. This system was introduced in 2016 in collaboration with Facebook to store rarely
accessed data cheaply for extended periods in data centers. The system was designed for BDXL
initially, then later configured to accommodate the AD discs.

Panasonic’s latest Data Archiver Writer contain optical units with typical transfer speeds of
54MB/s. However, the multi-unit writer is capable of simultaneously read from or write to 3
double-sided discs, or 6 disc-sides at a total of 324 MB/s.xv With LB-DH6 Data Archiver’s 2-
writer configuration, Freeze-ray data archivers can achieve a maximum transfer speed of 648 MB/s
per rack. The Panasonic Freeze-ray is shown and specified in Figure 56 and Table 10.

Figure 56. Panasonic’s Freeze-ray LB-DH6 (left) and LB-DH7 (right) data archiver systems. [47]
Reproduced with Permission: Panasonic Holdings Corporation

xv
Optical Disc Data Archiving: A New Age of Cold Data — The Storage Revolution Begins Now Whitepaper
(panasonic.net)

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Table 10. Freeze-ray specifications [48]

LB-DH6 Data Archiver LB-DH7 Data Archiver
Number of mountable magazines 532 532
Compatible Magazine Types 6TB (Gen 2 AD), 3TB (Gen 1 3TB (Gen 1 AD), 1.2TB (BD)
AD), 1.2TB (BD)
Capacity 3.19 PB 1.9 PB
Number of writer units (per rack) 2 unites 2 units
Total data transfer rate 648MB/s Read, 432MB/s Write 432 MB/s Read/Write
Host interface SAS/iSCSI/FC (by server for Data Archiver control software)
Command protocol SCSI (MMC, SMC)
Max Height when mounted in a 19- 46U 46U
inch rack with EIA panels
Input power DC +24V, +12V

TECHNICAL ISSUES
Current optical media solutions face an array of technical issues that are being addressed by the
industry. The commercially available products don’t offer any long-term development roadmap
that can keep up with the exponential growth of archival data needs.

CAPACITY
The latest Archival Disc (AD) features a per disc capacity of 500GB. A 6TB, 12-disc cartridge is
comparable in form factor with LTO tape cartridges and LFF HDDs. This capacity is considerably
less than HDDs and LTO tape cartridges. With the forecasted future capacity of 1TB/disc in the
next few years without further development plans published, current commercial optical storage
solutions will likely lag behind its magnetic counterparts.

COST
Table 6 lists the approximate costs of various data archiving media. It is apparent that current
optical technologies are not cost competitive from an up-front cost perspective. While lenient
environmental requirements and longer remastering cycle partially offset the media price
disadvantage, it still represents a large short-term commitment for enterprise clients to use optical
storage as the archival medium of choice.

SPEED
While read-write speed represents a challenge for nearly all digital storage technologies, a
particular challenge facing the optical enterprise storage solution is time-to-first-bit (TTFB). This
issue can be tackled by more efficient library systems that are able to retrieve and load discs into
multiple drives concurrently and using metadata to deliver initial data quickly. Even with current
TTFB, optical libraries deliver initial data much faster than tape.

POTENTIAL SOLUTIONS (ROADMAP OF QUANTIFIED KEY ATTRIBUTES

NEEDS)
The prevailing approach building on the multilayer concept, centers around storage in 3rd
dimension (and beyond) to maximize storage density both by increasing the areal density and the
number of layers. Concepts for wavelength multiplexing and multilevel data schemes are being

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

researched. In addition, concepts such as multiphoton writing and holographic storage are being
investigated.

Opportunities for all these approaches mentioned involve a pivot from the old, consumer-based
systems approach toward new concepts at the system level appropriate for data center deployment.
A particularly important example is the possibility of a system implementation that separates the
write hardware from the read hardware, a remote-write library. This approach, having read-only
drives, takes best advantage of the WORM nature, assuring an “air gap” for maximum cyber
security. In addition, it allows for using femtosecond laser technology for new multiphoton writing
mechanisms. In such remote write libraries, the high cost of femtosecond lasers can be shared and
amortized among, perhaps, millions of TB-scale media.

The high intensity of femtosecond lasers introduces new levels of multiphoton processes, such as
multiphoton ionization. This highly nonlinear response opens the door for many more potential
materials beyond the two-photon materials investigated years ago, with examples below. We note
that these highly nonlinear processes present the opportunity for data marks far below the
diffraction limit as only the central portion of the focal volume exceeds the writing threshold in
the nonlinear regime.

Additionally, as with any nonlinear writing process, absorption occurs only in the focal volume so
that materials transparent to the laser wavelength can be employed. The penetration and number
of written layers then become limited by the ability of optical system.

Historical and current optical discs are based on confocal imaging optics for both write and read.
The focused spot on the disc forms the data mark. For reading, the spot is imaged in a similar
confocal arrangement onto a segmented detector for implementing both focus and tracking servos
as well as the high-speed read channel. The time-dependent linear read channel is processed in a
similar manner to HDD read channels. Panasonic and Sony’s great improvements in the read
channel in developing the Archival Disc (as well as materials innovations) have resulted in bit
error rates as good as any media.

Advances in spatial light modulators (SLM) have provided the opportunity to focus an array of
writing spots all at once as described in the Piql section above. As described below, these devices
provide the opportunity for using a single objective lens for a high intensity write laser in a broader
array of materials. This has the potential to greatly increase the write speed as SLMs can have, at
least, thousands to millions of pixels. The writing threshold of the material and laser power will
limit how many marks can be written with a single laser pulse. The ability to store data at areal
densities using SLMs that approach or exceed the current state of the art for traditional high NA
writing, to our knowledge, has not yet been demonstrated.

On the read channel side, wide-field imaging can create a two- or three-dimensional images where
modern high-speed imaging and image processing techniques could interpret the data. For
example, machine learning has already been shown to provide high fidelity linear read channels in
HDDs and might be adapted to higher dimensions with multidimensional neural networks [49].
Read channels based on arrays of bits at once opens the door for bitwise writing schemes to enter
the page-wise read domain currently occupied by holographic storage.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

While these new write and read channel opportunities seem quite exciting, no performance data
for high-speed multi-dimensional write and read channels have been published, most notably, bit
error rates. In current technologies, exquisite servo controls are necessary to provide the high
fidelity write and read demanded by data storage and recovery and it is unclear what will be needed
in these new approaches.

FUTURE MULTILAYER TECHNOLOGY

The limitations of the manufacturing process used to make current blue laser discs has stimulated
intense development of higher areal density as in the Archival Disc. Folio Photonics Inc.
(www.foliophotonics.com) is adapting a widespread multilayer polymer manufacturing process to
both increase the number of layers and reduce costs [50]. The process is an adaptation of the co-
extrusion manufacturing process to produce active layers with ~100 nm thickness and separated
by ~10 micron thick buffer layers [51]. The process can produce dozens of layers “all at once” in
a roll-to-roll process (Figure 57). As the process is easier to scale up than down, many layers are
produced at much lower cost. Thus, the process promises to address both the cost and capacity
problems of existing technologies.

Melt Pump A
AB Feedblock

Extruder A Extruder B

Melt Pump B

Transfer Layer Multipliers

Tube
Skin Layer
Extruder Skin Layer
Feedblock
Melt Pump

Exit Die

Core layers

Figure 57. Folio Photonics manufacturing process and product. [52]

Reproduced with permission: Folio Photonics, Inc.

Both reflective and fluorescent data strategies have been demonstrated by a photothermal threshold
scheme similar to existing technology. The number of layers is not limited by the manufacturing
method, but by the laser power budget and spherical aberration correction in the objective lens of
the optical pickup unit.

Unlike current multilayer Blu-ray technology, the Folio Photonics disc does not have the spiral
tracking feature embossed on every layer. Such discs have been investigated by several
companies in the past and are known as super multilayer discs (Figure 58) [53].

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 58. Guide layer tracking scheme. [54]

In these cases, tracking is carried out by guide layer tracking where an additional laser is focused
on the disc substrate where the tracking pattern is embossed [55]. Folio’s scheme for creating a
focus error signal is shown in Figure 59 [56].

Figure 59. Scheme of Folio Photonic’s confocal fiber focus error signal scheme for fluorescent
signals. [57]
Reproduced with permission: Folio Photonics, Inc.

Folio Photonics has proved the feasibility of its optical pickup unit for writing and reading eight
layers at speed. The company believes that its unique roll-to-roll co-extrusion fabrication of
multilayer films, a process commonly used for other low-cost industry applications, enables the
company to produce discs for sale at $5/TB, making the solution to be cost-competitive with LTO
tapes and HDDs. With its first commercial product expected to launch in the general market in
2026 following testing and qualification, the company is in the process of creating a platform
ecosystem with library companies and other partners.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Folio’s current plans center on traditional confocal laser diode writing and linear read channels to
take advantage of existing supply chains and know-how. The future areal density roadmap includes
multilevel writing and wavelength multiplexing. Future implementations could take advantage of
the system level opportunities and new optical and image processing technologies described above.

Folio’s first product will be a two-sided 8-layer disc expected to hold from 800GB to 1TB of data,
with a roadmap to add additional layers over time as shown in Table 11.

OPTICAL MASS DATA STORAGE TECHNOLOGY ROADMAP

Table 11. Optical Mass Data Storage Technology Roadmap – BDXL, Archival Disc, and Folio
Photonic Disc
Technology Metrics Unit 2021 2023 2025 2027 2029 2031 2033 2035 2037
BDXL Capacity TB 0.2 0.2 No Further Roadmap Announced
Media Cost $/TB 40 40
Data Transfer Rate (Max) MB/sec 27 36
Layer Count (Double Side) Layers 6 6i
Areal Density (per Layer) GB/Layer 33 33

Archival Disc Capacity TB 0.5 0.5 1ii No Further Roadmap Announced

Media Cost $/TB 33 30 30iii
Data Transfer Rate (Max) MB/sec 375 375 750iv
Layer Count (Double Side) Layers 6 6 6
Areal Density (per Layer) GB/Layer 83 83 167

Folio Photonic Capacity TB 0.5-1.0 1.0-2.0 4 8 16 32 64

Disc
Media Cost $/TB 5 3 2 1 0.6 0.35 0.2
Data Transfer Rate (Max)v MB/sec 40 80 320 500 600 800 1000
Layer Count (Double Side) Layers 16 24 32 40 40 54 64
Areal Density (per Layer) GB/Layer 33 85 125 200 400 600 1000
i
BD-Double-Sided Disc (BD-DSD) has no plan to expand to quad-layer per side
ii
There is currently no set release date for the 3rd generation AD
iii
Assumes constant rate of cost/TB reduction relative to reduction from 300GB AD to 500GB AD
iv
Assumes no change in drive speed
v
Assumes multiple optical pickup units per drive
Note: Folio Photonic Disc will utilize its 2nd gen technology starting in 2031

EMERGING CONCEPTS
In this section, new concepts in optical data storage are summarized, namely, multiphoton writing
and holographic. These involve new twists on methods investigated in the last two decades. New
concepts include the use of ultrafast lasers in remote-write libraries, where WORM media are
written with specialized write-only drives that allow the high cost of ultrafast lasers to be amortized
among many thousands (or more) media. Group 47 and Cerabyte use 1- and 2-d spatial light
modulators to write many data marks at once and 2-d image analysis for reading. These concepts
have been commercialized by Piql for data preservation which have less rigorous requirements for

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

storage density and speed. Time will tell if these new concepts achieve the performance to replace
the traditional bit-wise disc technology in data center applications, where bit-wise methods also
develop apace.
MULTI-PHOTON
Project Silica (Glass)
A recent result at the University of Southampton has made it possible to store data in fused silica
(i.e., quartz glass) [58]. By focusing a femtosecond laser inside a block of fused silica, a 3D
nanostructure can be formed, which is a permanent change to the physical structure of the material
via multiphoton ionization.

Microsoft Research is collaborating with the University of Southampton to develop a cloud-scale

archiving system based on this technology [59]. The effort focuses on a design aimed at cloud-
scale performance by addressing such issues as the entanglement of the write and read throughputs,
the refresh cycle described above and constrained workloads by broadening the storage tier.

The use of silica has the potential benefit of true 3D optical storage, since the opacity of the quartz
glass is significantly less than the relatively opaque layers of a classical optical disc, the effect of
scattering and noise will be reduced, enabling multi-layer read/write. Due to the small nonlinear
optical response of silica, an amplified femtosecond laser is required for writing single bits. The
writing speed is thus given by the laser pulse repetition rate. Elongated marks exhibiting form
birefringence can be polarization multiplexed within a single voxel. In addition, multilevel storage
within a voxel has also been demonstrated. Together these are dubbed 5D optical storage. Seven
bits per voxel have been demonstrated. This multibit voxel storage leads to high storage densities.
This scheme is depicted in Figure 60.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 60. 5D storage of “The Hitchhiker’s Guide to the Galaxy.”

a) Illustration of data encoding and decoding. b) The birefringent images of data voxels of
different layers. Inset is the transmission of 100-layer data in the visible range. c) The birefringent
images after removing the background of (b). Insets are enlargements of small region (10 μm x 10
μm). d) Polar diagram of the measured retardance and azimuth of all voxels in (c). [60]
© Wang et al, Laser & Photonics Reviews published by Wiley-VCH GmbH. CC BY 4.0

The permanent nature of quartz glass also allows the realization of ultra-long-term storage, since
fused silica can retain its properties for hundreds of years while withstanding extreme conditions.
The nanograting in fused silica has high thermal stability and high optical damage threshold.
Project Silica indicates that its glass media can withstand temperatures over 1000 C, referring glass
storage as the modern day “stone etching”.

A recent video posted by Microsoft has revealed an interesting remote-write library system and
described their technology in some depth [61]. The system shelves the raw silica media on
passive storage panels. Silica media can be retrieved using shuttles, free-roaming robotics that
can be added or removed from the system based on changing archival access demand. The
library leverages an inherent air-gap design that eliminates the shuttles’ ability to move written
Silica media into writers. The decentralized and modular design of the library system mitigates
the risk of large-scale incidents and safeguards archival data integrity.

An early related publication as well as a recent one indicate a slow writing speed below
100kB/sec [62], [63]. However, the recent video suggests advances at the drive and system levels
have achieved “aggregate system level throughputs comparable to system-level tape archive
deployments.” We speculate that high-bandwidth spatial-light-modulators and high-rep rate
lasers are factors in this speed increase. The video also indicates volumetric storage densities
THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023
COPYRIGHT © 2023 IEEE. ALL RIGHTS RESERVED.
96

much higher than tape via 200-layer writes within the volume. For reading, machine learning is
used to decode a 2-D array of data using a convolutional neural network. We look forward to
learning more about the performance of the system.

Group 47 (DOTS)
A California-based start-up Group 47 (www.group47.com) has developed the Digital Optical Tape
System (DOTS) that preserves data for more than 200 years. DOTS leverages a legacy Kodak-
developed phase change media composed of a metallic alloy sputtered on a polyester-based film,
which is immune to electromagnetic fields (including EMP), chemically inert, and will not oxidize
and contains no chemical binders that can degrade hygroscopically. Data may be stored as binary
dots and blanks, as well human-eye readable images. DOTS has experimentally demonstrated a
200 to 2000 year archival longevity and a -9°C to 66°C temperature tolerance [64].

DOTS’s reflective encoding technique is based on the metallic alloy’s phase change properties.
Writing is a threshold photothermal process with a threshold of 0.3nJ/mm2. Both the reading and
writing processes are accelerated by novel paralleling and image encoding schemes [65]. The film
can be encoded and read continuously using the optical guide tracks and various guide spots, which
enables the medium to maintain data integrity and robustness against the effects of variable media
velocity and even physical deformation (Figure 61). DOTS’ write head uses a diffractive spatial
light modulator (SLM) as part of the laser multiplier that enables simultaneous writing of over
10,000 spots, allowing a single pass to fill up the tape. Writing is carried out using a 532 nm
wavelength laser. Lines of data are read by imaging the data using an oversampled linear detector
array.

While DOTS is yet to be commercialized, several successful prototype projects, including one
awarded by the Central Intelligence Agency (CIA), have demonstrated the solution successful
proof-of-concept. The library will use the remote write library concept. Group 47 also has plans to
release open source read-only units and license them to be manufactured royalty-free to accelerate
adoption. Group 47 indicates that commercial shipments in the next two years with a 1.2TB Native
capacity per tape cartridge and an impressive transfer speed of 1GB/s.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 61. Sample DOTS data encoding mechanism. [66]

Reproduced with permission: Group 47, Inc.

Cerabytexvi
Cerabyte (www.cerabyte.com), founded in 2020, is developing a patented [67] data storage
medium featuring a sputtered deposited extremely durable ceramic nano layer (10 nm thick) with
a broad absorption spectrum, a “grey ceramic”, which allows ultra-fast write with threshold as low
as 0.1 nJ per bit. The ceramic is deposited on both sides of a flexible ultra-thin planar substrate
with only 100 µm thick foldable glass [68] or 10 µm thick ribbon glass [69] for potential tape
development. These substrates and the coating of a ceramic nano layers leverage existing display
glass production capacities of today 350 million m2 per year [70]. Thus, the company expects to
achieve media cost below $ 1 per TB by 2030.

Accelerated aging tests at temperatures of -273 °C to 500 °C indicate potential storage lifetime in
the millennia range. Furthermore, the data is not corrupted even when exposed to electromagnetic
pulses, UV and gamma rays.

Data is written encoded in an array of data matrices by using a 2-D digital micro mirror device
(DMD) with up to 2 million elements simultaneously written by femtosecond laser pulses in the
UV spectrum with a rep rate of several kHz (Figure 62). This implies a writing speed in excess of
1 GB/s with less than 1 W average power.

xvi
This section authored by Cerabyte (Ceramic Data Solutions Holding GmbH) CEO, Christian Pflaum

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

High- speed laser writing with DMD Microscope optic for writing & reading High- speed reading via image sensor

Illumination for
reading

Fem r
Las
DMD Microscope Auto

t os
e
optic focus

ec o
ed
Hig h sp e or
ns

nd
Write Read im age se

XY stage

Figure 62. Cerabyte writing scheme

Reading is performed at GB/sec rate using high speed image sensors, UV illumination and parallel,
high speed image processing for decoding. The high-resolution images are captured at more than
500 fps followed by parallel processed by a FPGA, which produces a data stream using the 2-D
image while applying conventional error correction methods in a second processing step. Both
reading and writing is carried out across the substrate by scanning the microscope optics using
high-speed XY stages kept in focus using a piezo driven auto focus system. This setup enables
random access.

Plans call for hundreds of 9 by 9 cm media sheets to be stacked in individual cartridges to reduce
storage volume as shown in Figure 63. Cerabyte uses the form factor of mainstream magnetic tape
cartridges. The media access scheme enables random access enabling a faster time to first bit than
tape.

Storage media stacked in cartridge Ceramic data storage medium Data coded in matrix format

Figure 63. Cerabyte’s storage concept

Cerabyte employs a commercially available library unit utilizing a remote write architecture. The
library will locate and retrieve the cartridge, then unload and unstack the substrates for positioning
the addressed substrate in the optical unit.

Cerabyte indicates that a demo system with a single write & read head unit achieving 100 MB/s
write/read speeds and a storage capacity of 1PB per 19” rack will be developed in 2023. The first
product for corporate archiving systems is scheduled to be launched in 2024 with 500 MB/s
write/read speeds and a capacity of 5 PB/rack up to ten 19” racks. In 2025 a 10-30PB rack system
for cloud data centers will be launched with 1 GB/s+ write/read speeds, which is projected to

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

increase over time (Figure 64) and will achieve capacities and bandwidth attractive for hyperscaler
use cases by the end of the decade.

2023 2024 2025- 27 2028- 30

Demo On- Prem Cloud Data Hyperscaler

System Data Center Center Data Center
2023

19" 19" 19" 19"

1PB/rack 5 PB/rack 10- 30 PB/rack 60- 100 PB/rack
100 MB/s 500 MB/s 1+ GB/s 2+ GB/s
<90 sec to first bit <30 sec to first bit <15 sec to first bit <10 sec to first bit

Figure 64. Cerabyte’s product roadmap of enterprise archiving systems

HOLOGRAPHIC
While holographic data storage was proposed as early as the 1960s. By the 2000s, leading research
institutions including Aprilis and Bell Labs, have failed to deliver commercial success despite
significant research progress. Brief commercial attempts such as one by InPhase Technologies,
which released a $180, 300GB holographic storage disc (HSD) in 2007 [71], all failed to capture
the market due to their high costs [72].

The technology provides a series of benefits, including high data density, massively parallel write
and read, fast access time of <50 us, fast transfer rate of ~ 1 Gbit/s, air gap and WORM or
rewritable formats [73]. Despite these technological advantages, HSD development remains in the
R&D stage today, decades after the technology’s introduction. While many successful prototypes
have been built, none have gained significant commercial traction.

Holographic data storage (HDS) allows multiple holograms to be written in the same fixed volume
of rather thick media (200+ um), with theoretical recording density limit on the order of tens of
Tb/cm3. Data recorded holographically are read back in “pages” instead of bit streams. HDS is a
volumetric technique, making its density proportional to 1/l3, while for optical discs the density is
proportional to 1/l2. The storage density is a function of the number of holograms multiplexed into
a volume of the recording media and principally determined by its refractive index contrast and
media thickness.

A wide range of materials was investigated as holographic media, including photorefractive

crystals, photorefractive organic materials, photo-addressable polymers, photochrome systems,
photopolymers, systems with electrocyclic ring closure, etc. Multiple materials properties such as
dimensional stability (resistance of the material to shrinkage or expansion during the recording
process and thermal changes), optical quality, scattering level, linearity and volatility affect the
fidelity of the recording and read-out process. The recording and read-out transfer rates depend on
the photosensitivity and the diffraction efficiency.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

For the WORM media format, photopolymers are preferred materials that can be easily
manufactured in production scale. However, their shrinkage during writing is a critical issue that
should be solved before industrial applications is possible.

A lot of research efforts were made for developing writing and reading holographic systems.
Several multiplexing technologies were proposed the “page-wise” approach, including two-beam
angle multiplexing [74], coaxial multiplexing [75], and coaxial shift multiplexing [76]. Using these
methods several prototypes were built to demonstrate drive system capabilities to handle
dimensional changes of the materials due to thermal expansion.

PROJECT HSD (HOLOGRAPHIC)

Announced in 2020, as one of the two cloud-focused storage projects led by Microsoft Research,
Project HSD explores the potential of Holographic storage, designing mechanical-movement-free,
high-endurance cloud storage that is both performant and cost-effective [77], [78].

Holographic data formats can be similar to that of 2-dimensional QR codes, coded inside a LiNbO3
crystal. By leveraging two beams: the data beam and the reference beam, the 2-D codes can be
encoded as interference patterns. By changing the angle of the reference beam, one can record
multiple sets of interference patterns within the same block of crystal. Data can then be read using
the reference beam that recreates the interference pattern, before capturing the 2-D code using a
camera. The crystal can be then erased using UV light and be reused indefinitely to store more
data.

On the physical level, the Iron ions doped in the crystal add an additional donor level and a deep
trap state to the energy levels of the crystal. When a region is exposed to the bright part of the
interference pattern, the extra electrons are excited, moving to the conduction band, before
decaying preferentially to the deep iron trap level [79].

Their focus on scalable optical systems, storage systems and machine learning are important new
directions that can move holographic enterprise optical archiving forward. We speculate that
Project HSD’s approach addresses previous challenges.

CHALLENGES (CRITICAL ISSUES)

Currently, most of the enterprise archival market is dominated by LTO magnetic tape and HDD.
Magnetic tapes arguably offer the lowest cost/GB and a relatively long lifetime, positioning them
to be an effective solution for “deep archival” use cases. However, its long TTFB and the lack of
random access prevents it from delivering value in more active archival applications. LTO also
suffers from a relatively short lifetime and an expensive and not always backward-read-compatible
drive. HDD offers performance that exceeds the need of archival applications but suffers from a
high operating cost and short lifespan, which results in frequent remastering, further increasing
TCO. The challenges for optical storage media to compete with magnetic tape and HDD in the
near to medium term can be categorized into three aspects – capacity, cost, and read/write speed.

CAPACITY
While capacity technically can be expressed as a function of cost, the value of a higher GB/volume
is self-explanatory. High data density can be an insurmountable requirement for large data center

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

deployments as space and other operating constraints are often pre-determined factors regardless
of budget.

Layered optical discs traditionally drive capacity growth by increasing both the number of layers
per disc and areal density. However, spherical aberrations and light-attenuation would eventually
limit the number of layers that can be efficiently recorded and retrieved. However, multiphoton
writing/reading demonstrations can increase volumetric and areal density. In addition, 3D
approaches such as multiphoton in non-layered media and holographic storage promise high
volumetric density.

COST
The capacity challenge can be addressed as an aspect of the broad cost challenges. LTO-8 tapes
are priced at around $5/TB, which is an order of magnitude cheaper than the optical competitors,
including the BDXL and Sony-Panasonic Archival discs. Optical disc is not competitive compared
to the much more expensive HDD either, despite the slowing of HDD capacity improvement and
cost reduction. With a shrinking market, optical disc manufacturers are benefitting less from the
economy of scale. To overcome this challenge, optical media must move to reduce the cost in
media manufacturing that utilizes more efficient processes, lower-cost materials, and/or other cost-
cutting innovations. System-level costs for enterprise data storage including drives and libraries
should be competitive with the competition with the caveat that optical data storage approaches
have an intrinsically lower total cost of ownership owing to their long life, high energy efficiency,
and low carbon footprint.

SPEED
The known approaches to optical data storage involves libraries of platters and spooled media.
Platter libraries have a much lower time to first bit of any spooled media, likely in the seconds to
tens of seconds. While the read/write speed has been a challenge for most storage media due to the
rapid growth of data size, drive and system-level innovations have largely overcome this hurdle
on the enterprise scale by implementing multi-channel approaches such as storing across multiple
drives within a disc library, or as Sony has, by creating a drive with multiple pickup units. Machine
learning using neural networks for reading has the potential to substantially increase reading speed
by reading in multiple dimensions. In addition, recent results using machine learning in the data
channel of hard drives show considerable promise improved performance [80], [81]. Page-wise
reading in holographic storage would also result in high throughput.

SUMMARY & KEY POINTS

• Optical storage media is pivoting from the historical consumer media distribution focus to
the enterprise/institutional archival storage use case.
• To expand capacity while reducing cost, new optical technologies will continue to exploit
the 3rd dimension and beyond.
• Robust library systems designed for optical media are being engineered to meet the robust
requirements of enterprise-level storage needs.
• Optical media’s low maintenance/operating energy cost and remastering frequency will
realize a natural advantage in the sustainability-conscious data archive landscape.
• Optical WORM technologies the air-gap offer advantages in cybersecurity.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

• Optical technologies present the lowest energy consumption, both intrinsically and in data
center environment control.
• Optical technologies promise significant greenhouse gas reduction.
• Critical challenges facing optical data storage are lowering initial cost, expanding capacity,
increasing speed, and error management.
• Non-disc, novel technologies will likely deliver breakthroughs in capacity, cost, and speed
to the data storage industry at-large, though likely not in the near-term.
• The possibility of remote-write libraries, femtosecond lasers and high-speed display and
imaging technologies provide new opportunities for enterprise optical data storage.

CONCLUSIONS & RECOMMENDATIONS

The demand for archival data storage is growing rapidly due to the growth of permanent access
driven by demands of data science, internet of things, and requirement for permanent storage. At
the same time, the pace of capacity growth and cost reduction in incumbent magnetic technologies
are slowing due to the approach of physical limits implied by Kryder’s law. In addition, the
oligopoly of both consumers and producers are stressing the data storage market. Optical archiving
technologies are now emerging based on multilayer disc libraries, but still suffer from high cost
and low capacity. New approaches to multilayer manufacturing aimed at lowering cost and
increasing capacity are in development, while other approaches based on holographic storage and
multiphoton writing are being researched.

For too many years, R&D in the development of optical data storage has focused too much on
media capacity while downplaying the importance of speed and cost. In the future, the challenges
of developing drives and libraries need to be considered from day one. Partnerships among media,
drive, and library system players within and between organizations will be key to widespread
commercialization of optical archiving technologies and the strong value proposition they provide.

REFERENCES

1
“IFPI GLOBAL MUSIC REPORT 2023”. IFPI. https://globalmusicreport.ifpi.org/ (accessed September 1st.
2023)
2
“IFPI GLOBAL MUSIC REPORT 2023”. IFPI. https://globalmusicreport.ifpi.org/ (accessed September 1st.
2023) web.archive.org/web/20150518082128/http:/5eshs.hpdst.gr/abstracts/139""The history of ideas "the
optical disc as a "unique" carrier of information in the systems management". European Society of the History
of Science. Archived from the original on 2015-05-18.
3
"The history of ideas "the optical disc as a "unique" carrier of information in the systems management".
European Society of the History of Science. Archived from the original on 2015-05-18.
4
E. Hwu & A. Boisen, “Hacking CD/DVD/Blu-ray for Biosensing,” ACS Sens. 2018, 3, 7, 1222–1232, July
2018, doi: 10.1021/acssensors.8b00340.
5
“U.S. Sales Database.” RIAA. https://www.riaa.com/u-s-sales-database/ (accessed December 1st, 2022)
6
S. Whitten, “The death of the DVD: Why sales dropped more than 86% in 13 years.” CNBC.
https://www.cnbc.com/2019/11/08/the-death-of-the-dvd-why-sales-dropped-more-than-86percent-in-13-
years.html (accessed December 1st, 2022)

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

7
M. Siegler, “Yep, Apple Killed The CD Today.” TechCrunch. https://techcrunch.com/2010/10/20/a-compact-
death/ (accessed December 1st, 2022)
8
N. Statt, “Sony announces PlayStation 5 Digital Edition with no disc drive.” The Verge.
https://www.theverge.com/2020/6/11/21288493/ps5-playstation-5-digital-edition-no-disc-drive-hardware-specs-
price-sony (accessed December 1st, 2022)
9
“U.S. Sales Database.” RIAA. https://www.riaa.com/u-s-sales-database/ (accessed December 1st, 2022)
10
S. Whitten, “The death of the DVD: Why sales dropped more than 86% in 13 years.” CNBC.
https://www.cnbc.com/2019/11/08/the-death-of-the-dvd-why-sales-dropped-more-than-86percent-in-13-
years.html (accessed December 1st, 2022)
11
“ECMA-167.” Ecma International. https://www.ecma-international.org/publications-and-
standards/standards/ecma-167/ (accessed September 1st, 2023)
12
2019 iNEMI Roadmap - Mass Data Storage. iNEMI. (accessed December 1st, 2022)
13
C. Walter, “Kryder's Law: The doubling of processor speed every 18 months is a snail's pace compared with
rising hard-disk capacity, and Mark Kryder plans to squeeze in even more bits.”
https://www.scientificamerican.com/article/kryders-law/ (accessed September 1st, 2023)
14
“Optical Technology: Optical Data Archiver freeze-ray series.” Panasonic Global.
https://panasonic.net/cns/archiver/optical_technology/index.html (accessed December 1st, 2022)
15
“Optical Technology: Optical Data Archiver freeze-ray series.” Panasonic Global.
https://panasonic.net/cns/archiver/optical_technology/index.html (accessed December 1st, 2022)
16
“White Paper: Archival Disc Technology.” 1st Edition. Panasonic.net. July 2015
https://panasonic.net/cns/archiver/pdf/E_WhitePaper_ArchivalDisc_Ver100.pdf (accessed December 1st, 2022)
17
“White Paper: Archival Disc Technology.” 2nd Edition. Panasonic.net. January 2020
https://panasonic.net/cns/archiver/pdf/E_WhitePaper_ArchivalDisc_2nd_Edition.pdf (accessed December 1st,
2022)
18
“White Paper: Archival Disc Technology.” 1st Edition. Panasonic.net. July 2015
https://panasonic.net/cns/archiver/pdf/E_WhitePaper_ArchivalDisc_Ver100.pdf (accessed December 1st, 2022)
19
“White Paper: Archival Disc Technology.” 2nd Edition. Panasonic.net. January 2020
https://panasonic.net/cns/archiver/pdf/E_WhitePaper_ArchivalDisc_2nd_Edition.pdf (accessed December 1st,
2022)
20
J. Moore, “The Escalating Challenge of Preserving Enterprise Data.” Further Market Research. August 2022
https://asset.fujifilm.com/master/americas/files/2022-
08/5082fc03fc44d80b691c87a1a96febd5/Furthur_Market_Research_WP_080322_FINAL.pdf (accessed
December 1st, 2022)
21
“White Paper: Blu-ray Disc Format - 1. B Physical Format Specifications for BD-R.” 5th Edition. October 2010
https://blog.ligos.net/images/The-Reliability-Of-Optical-Disks/BD-R_physical_specifications-18326.pdf
(accessed December 1st, 2022)
22
“Format Specification - R3 Format Specification (BDXL™).” blu-raydisc.info. https://blu-raydisc.info/format-
spec/r3-spec.php (accessed December 1st, 2022)
23
“Blu-ray Disc Association Announces New Double-Sided Disc Specification for ‘Big Data’ Storage.” blu-
raydisc.com. 20140818-Blu-ray-Disc-Association-Announces-New-Double-Sided-Disc-Specification-for-Big-
Data-Storage.pdf (blu-raydisc.com) (accessed December 1st, 2022)
24
P. Lilly, “Pioneer Just Made A New Optical Disc Drive For PCs And Yes It's 2022.” HotHardware.
https://hothardware.com/news/pioneer-optical-disc-drive-pc (accessed December 1st, 2022)
25
“White Paper: Blu-ray Disc Format - 1. B Physical Format Specifications for BD-R.” 5th Edition. October 2010
https://blog.ligos.net/images/The-Reliability-Of-Optical-Disks/BD-R_physical_specifications-18326.pdf
(accessed December 1st, 2022)

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

26
“White Paper: Archival Disc Technology.” 1st Edition. Panasonic.net. July 2015
https://panasonic.net/cns/archiver/pdf/E_WhitePaper_ArchivalDisc_Ver100.pdf (accessed December 1st, 2022)
27
“White Paper: Archival Disc Technology.” 2nd Edition. Panasonic.net. January 2020
https://panasonic.net/cns/archiver/pdf/E_WhitePaper_ArchivalDisc_2nd_Edition.pdf (accessed December 1st,
2022)
28
“White Paper: Archival Disc Technology.” 2nd Edition. Panasonic.net. January 2020
https://panasonic.net/cns/archiver/pdf/E_WhitePaper_ArchivalDisc_2nd_Edition.pdf (accessed December 1st,
2022)
29
“White Paper: Archival Disc Technology.” 1st Edition. Panasonic.net. July 2015
https://panasonic.net/cns/archiver/pdf/E_WhitePaper_ArchivalDisc_Ver100.pdf (accessed December 1st, 2022)
30
“White Paper: Archival Disc Technology.” 2nd Edition. Panasonic.net. January 2020
https://panasonic.net/cns/archiver/pdf/E_WhitePaper_ArchivalDisc_2nd_Edition.pdf (accessed December 1st,
2022)
31
“Long-term Data Storage.” piql. https://piql.com/services/long-term-data-storage/ (accessed September 1st,
2023
32
“Technology.” piql. https://piql.com/about/technology/ (accessed September 1st, 2023)
33
M. Smith, “How To Back Up Your Digital Photos... To Film.” Fstoppers. https://fstoppers.com/originals/how-
back-your-digital-photos-film-537562 (accessed December 1st, 2022)
34
“Arctic World Archive.” Wikipedia. https://en.wikipedia.org/wiki/Arctic_World_Archive (accessed September
1st, 2023)
35
“Suzhou NETZON Information Storage Technology.” hit-netzon.com.cn. https://www.hit-
netzon.com.cn/HDL10368-en.html (accessed December 1st, 2022)
36
“Suzhou NETZON Information Storage Technology.” hit-netzon.com.cn. https://www.hit-
netzon.com.cn/HDL10368-en.html (accessed December 1st, 2022)
37
“Commercial Products.” Amethystum.com. https://www.amethystum.com/article/10171.html (accessed
December 1st, 2022)
38
“Commercial Products.” Amethystum.com. https://www.amethystum.com/article/10171.html (accessed
December 1st, 2022)
39
“Zerras.” ICEBOX. https://www.zerras.com/icebox/ (accessed December 1st, 2022)
40
“Disc Library Systems.” Kintronics. https://kintronics.com/disc-library-systems/ (accessed December 1st, 2022)
41
M. Smolaks, “Optical disc as the future of data archiving.” DCD.
https://www.datacenterdynamics.com/en/analysis/optical-disc-as-the-future-of-data-archiving/ (accessed
December 1st, 2022)
42
“Optical Disc Archive Generation 3.” Sony Pro.
https://pro.sony/s3/2020/05/28164824/Sony_ODA_Generation3_Brochure-PDF.pdf (accessed September 1st,
2023)
43
“Optical Disc Archive Generation 3.” Sony Pro.
https://pro.sony/s3/2020/05/28164824/Sony_ODA_Generation3_Brochure-PDF.pdf (accessed September 1st,
2023)
44
“Optical Disc Archive Generation 3.” Sony Pro.
https://pro.sony/s3/2020/05/28164824/Sony_ODA_Generation3_Brochure-PDF.pdf (accessed September 1st,
2023)
45
“Cold Data Storage Technology.” Sony Pro. https://pro.sony/ue_US/technology/optical-disc-archive (accessed
December 1st, 2022)

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

46
“光ディスクデータアーカイブシステム - パナソニックコネクト” panasonic.com. (accessed September
1st, 2023); “Products | Optical Data Archiver freeze-ray series.” Panasonic Global.
https://panasonic.net/cns/archiver/product/#Outline (accessed December 1st, 2022)
47
“光ディスクデータアーカイブシステム - パナソニックコネクト” panasonic.com. (accessed September
1st, 2023); “Products | Optical Data Archiver freeze-ray series.” Panasonic Global.
https://panasonic.net/cns/archiver/product/#Outline (accessed December 1st, 2022)
48
“光ディスクデータアーカイブシステム - パナソニックコネクト” panasonic.com. (accessed September
1st, 2023); “Products | Optical Data Archiver freeze-ray series.” Panasonic Global.
https://panasonic.net/cns/archiver/product/#Outline (accessed December 1st, 2022)
49
Y. Qin and J. -G. Zhu, "Automatically Resolving Intertrack Interference With Convolution Neural Network
Detection Channel in TDMR," IEEE Transactions on Magnetics, vol. 57, no. 2, pp. 1-6, Feb. 2021, Art no.
3100306, doi: 10.1109/TMAG.2020.3017373.
50
Folio Photonics, Inc.
51
C. Mellor, “Folio Photonics archive optical disk technology is real.” Blocks and Files.
https://blocksandfiles.com/2022/08/31/folio-photonics-archive-optical-disk-technology-is-real/ (accessed
December 1st, 2022)
52
Folio Photonics, Inc.
53
Y. Tanaka, T. Ogata, & S. Imagawa, “Decoupled direct tracking control system based on use of a virtual track
for multilayer disk with a separate guide layer,” Jpn. J. Appl. Phys. 54 09MB03, August 2015, doi:
10.7567/JJAP.54.09MB03.
54
Y. Tanaka, T. Ogata, & S. Imagawa, “Decoupled direct tracking control system based on use of a virtual track
for multilayer disk with a separate guide layer,” Jpn. J. Appl. Phys. 54 09MB03, August 2015, doi:
10.7567/JJAP.54.09MB03.
55
M. Ogasawara, K. Takahashi, M. Nakano, M. Inoue, A. Kosuda, and T. Kikukawa, “Sixteen-Layer Write Once
Disc with a Separated Guide Layer,” Jpn. J. Appl. Phys. 50 09MF01, September 2011, doi:
10.1143/JJAP.50.09MF01.
56
Patent: US 11,456,010 B2: SYSTEMS AND METHODS FOR INCREASING DATA RATE AND STORAGE
DENSITY IN MULTILAYER OPTICAL DISCS
57
Patent: US 11,456,010 B2: SYSTEMS AND METHODS FOR INCREASING DATA RATE AND STORAGE
DENSITY IN MULTILAYER OPTICAL DISCS
58
“Project Silica.” Microsoft. https://www.microsoft.com/en-us/research/project/project-silica/ (accessed
September 1st, 2023)
59
P. Anderson et al., “Glass: A New Media for a New Era?” 10th USENIX Workshop on Hot Topics in Storage
and File Systems (HotStorage 18) July 2018.
60
H. Wang et al., “100‐Layer Error‐Free 5D Optical Data Storage by Ultrafast Laser Nanostructuring in Glass,”
Laser & Photonics Reviews, 16(4), pp. 2100563, January 2022, doi: 10.1002/lpor.202100563.
61
“Research talk: Storing data for millennia.” Microsoft Research. Youtube. https://youtu.be/V7L_wdEuQXs
(accessed September 1st, 2023)
62
J. Zhang et al., “Seemingly Unlimited Lifetime Data Storage in Nanostructured Glass,” Phys. Rev. Lett. 112,
033901, January 2014, doi: 10.1103/PhysRevLett.112.033901.
63
A. Jain et al., “Optimization of Multi-Layer Data Recording and Reading in an Optical Disc,” Photonics 9(10),
690. September 2022, doi: 10.3390/photonics9100690.
64
Group 47
65
US Patents: US 9,208,813 B2, US 9,508,376 B2, US 9,640,214 B2, US 10,033,961 B2, US 10,067,697 B2
66
Group 47

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

67
Patents: WO2021028035 Ceramic data storage media; WO2022033800 Data recording on ceramic data storage
media: WO2022194354 Super resolution reading of ceramic data storage media
68
“Foldable Glass | Ultra-thin Bendable Glass Technology.” Corning.
https://www.corning.com/worldwide/en/innovation/materials-science/glass/Foldable-Glass.html (accessed
September 1st, 2023)
69
“Glass-ribbon.” Nippon Electric Glass Co., Ltd. https://www.neg.co.jp/en/product/ep/glass-ribbon (accessed
September 1st, 2023)
70
“Display production capacity by type 2016-2025.” Statista. https://www.statista.com/statistics/1057441/display-
panel-production-capacity-type/ (accessed September 1st, 2023)
71
P. Miller, “InPhase 300GB holographic storage solution out the door.” Engadget.
https://www.engadget.com/2007-02-13-inphase-300gb-holographic-storage-solution-out-the-door.html
(accessed September 1st, 2023)
72
P. Shread, “Holographic Storage Appears.” Enterprise Storage Forum.
https://www.enterprisestorageforum.com/hardware/holographic-storage-appears/ (accessed September 1st,
2023)
73
S. S. Orlov, E. Bjornson, W. Phillips, Y. Takashima, X. Li and L. Hesselink, "High transfer rate (1 Gbit/sec)
high-capacity holographic disk digital data storage system," Conference on Lasers and Electro-Optics (CLEO
2000). Technical Digest. Postconference Edition. TOPS Vol.39 (IEEE Cat. No.00CH37088), San Francisco,
CA, USA, 2000, pp. 190-191, doi: 10.1109/CLEO.2000.906896.
74
“A matter of format,” Nature Photon 2, 401, July 2008, doi: 10.1038/nphoton.2008.118.
75
P.Koppa et al., Multiphoton and Light Driven Multielectron Processes in Organics: New Phenomena, Materials
and Applications (Eds.: F. Kajzar, M.V. Agranovich), Kluwer, Dordrecht, 2000
76
K. Tanaka et al., “High Density Recording of 270 Gbit/in.2 in a Coaxial Holographic Recording System,” Jpn.
J. Appl. Phys. 47 5891, July 2008, doi: 10.1143/JJAP.47.5891.
77
“How does holographic storage work?” Microsoft Research. YouTube.
https://www.youtube.com/watch?v=4EADwGV5Gv8&t=20s&ab_channel=MicrosoftResearch (accessed
December 1st, 2022)
78
“Project HSD: Holographic Storage Device for the Cloud.” Microsoft Research. https://www.microsoft.com/en-
us/research/project/hsd/ (accessed December 1st, 2022)
79
“The physics of hologram formation in iron doped lithium niobate.” Microsoft Research. YouTube.
https://www.youtube.com/watch?v=CfME3e7aSNk&t=2s (accessed December 1st, 2022)
80
A. Aboutaleb et al., "Deep Neural Network-Based Detection and Partial Response Equalization for Multilayer
Magnetic Recording," IEEE Transactions on Magnetics, vol. 57, no. 3, pp. 1-12, March 2021, Art no. 3101012,
doi: 10.1109/TMAG.2020.3038435.
81
Y. Qin and J. -G. Zhu, "Automatically Resolving Intertrack Interference With Convolution Neural Network
Detection Channel in TDMR," IEEE Transactions on Magnetics, vol. 57, no. 2, pp. 1-6, Feb. 2021, Art no.
3100306, doi: 10.1109/TMAG.2020.3017373.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

DNA Data Storage

SITUATION ANALYSIS
The storage technologies outlined in the other chapters of this roadmap will continue to evolve and
scale, but the reality is that the world is attempting to digitize unprecedented amounts of
information. This information can be valuable if mined, stitched together, or otherwise searched
and analyzed; however, the cost of storing the massive amount of data associated with this
information is beginning to overwhelm the ability to pay for it using conventional storage
technologies, and a cost-effective scaling path using traditional storage technologies remains
unclear.

Further, the operational costs of refreshing data, or creating copies, using existing storage
technologies is becoming prohibitive, with the refresh of some large archives needing to start, or
nearly so, by the time the previous refresh finishes. These factors are, in turn, causing potentially
valuable data, and even the potential for newly discovered knowledge, to be thrown away. This
lost opportunity is shown as the “Zone of Potential Insufficiency” (Figure 65).

Figure 65. The Zone of Potential Insufficiency1

All of these factors are leading system designers to look for new storage technologies which can
sustain the capacity, access flexibility, and TCO needed for the massive wave of digitization, and
one of the leading technologies being considered is synthetic DNA.i

i
For background on DNA data storage, see two presentations from the “DNA Memories” tutorial at the 2023 IEEE
15th International Memory workshop (IMW 2023): 1) “Integrating biomolecules and semiconductors to build a data
storage system”, Andres Fernandez, Twist Bioscience; and 2) “DNA Sequencing for Data Storage”, Boyan
Boyanov, Illumina. Also see “Preserving our Digital Legacy: An Introduction to DNA Storage, DNA Data Storage
Alliance.”

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

DNA (Figure 66) is a potentially compelling storage medium due to its ~1bit/nm3 bit size
(Figure 67), the fact that it is very stable at room temperature if kept dry, thus potentially avoiding
fixity checks and media technology migration, and that the medium is decoupled from the
read/write device, meaning it can always be read. This combination of properties enables the
potential of the proverbial “datacenter in a shoebox”; in other words, compelling TCO.

Figure 66. DNA Double Helix: National Human Genome Research Institute (NHGRI)

Figure 67. DNA Density; Preserving our Digital Legacy: An Introduction to DNA Data Storage; DNA
Data Storage Alliance4

While an area of active research and development, DNA data storage (DDS) is not generally
commercially viable today and thus does not lend itself to a clear roadmap like the other
technologies documented in this report. For example, it is unclear what price/performance points
will drive broad commercialization of DDS, since the potential use cases are highly dependent on
techniques for writing (synthesis) and reading (sequencing) digital data in synthetic DNA which,
while well-grounded in decades of medical/scientific applications, are nascent in the context of
data storage.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

This means that, in this nascent phase of the DDS ecosystem, not only will synthesis and
sequencing capabilities continue to be tailored for specific use cases, the business value of any
solution (i.e., the price users are willing to pay) will be significantly affected by the capabilities
of the technology as it is deployed in production systems. Notwithstanding, this chapter outlines
the main challenges and issues related to the state of the DDS ecosystem today.

CHALLENGES AND ISSUES

TCO
Storage TCO encompasses all capital expenditures (Capex) and operational expenditures (Opex)
over the lifetime of the storage solution. Capex includes costs of equipment/hardware, software
and infrastructure. Opex includes consumables, labor, energy and natural resources (e.g., water),
and maintenance and support. Another important contribution to both Capex and Opex is data
migration, either when media reaches end of life, or when the capabilities of new media justify
transferring data.

As seen in recent industry analyses,2,3 tape offers significant TCO advantages over other
magnetic media which, when combined with performance, make it the dominant technology in
the archival tier.

Comparing projected costs of DDS to existing technologies like tape is not that useful because
the use case for a medium like DNA, especially initially, will not involve replacing tape, or any
other existing storage technologies, but complementing them in the storage hierarchy (Figure
68).

Moreover, the relationship of the value of a DDS solution to a particular use case will evolve
dynamically as DDS technology emerges commercially. The first use cases of DDS will likely be
for deep/cold archives (100+ years duration), with Write Once/Read Never-to-Seldom access
patterns, and assuming no additions or changes to the data during storage. For customers with such
use cases, the value of the data may outweigh a high (relative to incumbent media) cost of writing
and reading, as well as (again relative to incumbent media) low throughput and high latency (time
to first byte).

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 68. The Evolving Storage Pyramid4

Source: Gartner Product Manager Insight: Tape to the Future? October 23, 2020ii

As DNA write and read technologies improve in cost and performance, and are realized in more
working systems, the solutions will enable more flexible use cases, with more diverse access
patterns, resulting in different tradeoffs between read/write speed and cost versus the value of the
data (i.e., what is an acceptable TCO). Aside from increases in throughput and reduced latency,
the evolution of these TCO tradeoffs will involve many other aspects of using DNA as a storage
medium in a system, including:

a) Costs of write and read substrates and storage containers, their reusability, the costs of
reagents, resource costs (e.g., labor, water, electricity).
b) Cost of rack or system-level components necessary to automate storage and handling of
substrates, containers and/or reagents.
c) Transition from manual to automated instruments.
d) Optimized error rates (i.e., reduced number of molecules required per TB); and
e) Standardized interfaces between each step of the DDS workflow.

Lastly, all of the above will affect, and be affected by, how DNA data storage is delivered. For
example, writing and reading may be services, or on premises at a data center, or both.

By far the most definitive and important thing we can say about DDS TCO today is that when
DNA based media is stored in controlled environments (temperature and humidity), it is extremely
stable. Figure 69 shows DNA preservation methods studied in Organick5, Grass6, Coudy7, and
Bonnet8 that demonstrate that properly controlled preservation techniques can ensure molecular
stability (i.e., data stability), even at room temperature, over very long periods. Grass, Organick,

ii
Gartner Product Manager Insight: “Tape to the Future?” J. Monroe and R. Preston. 23 October 2020, ID
G000724101.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

and Coudy additionally showed that the observed levels of molecular stability were sufficient to
enable the encoded data to be successfully recovered.

Figure 69. Half-life of verious DNA preservation methods. 5

Source: From Figure 2b, Organick et al

The demonstrated stability of DNA can effectively eliminate fixity checks and technology
migration, both increasingly large factors in archival storage TCO. In particular, regarding
technology migration, since DNA based data is stored in a pool, and not written into a device
with a fixed substrate as media (i.e., HDD, SSD, Tape), and the DNA molecular structure is
universal, it will always be physically possible to read DNA media back from an archive, even if
the devices used to write or read the media when it was created are no longer available. This
means that DNA based data need not undergo the increasingly costly technology migration
which characterizes today’s archival storage technologies.

Figure 70 illustrates the implications of DNA data storage cost over time. The commercial price
to store a petabyte of data using list prices for Tape (Fujifilm Calculator) and Cloud (AWS
public pricing) was used as a baseline. The price points for DNA-based storage were picked
arbitrarily, for illustration. Beyond 10 years, the advantageous properties of DNA-based storage
relative to incumbent media begin to dominate; that is, the DNA data storage TCO becomes
increasingly attractive.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Figure 70. Estimated Cost of Writing and Storing – Legacy vs. DNA4
The key challenges for DNA data storage TCO are not related to storage, but to read/write
performance and cost. We explore this in the following two sections.

WRITING DNA (SYNTHESIS)

The main challenge for DDS in terms of writing is that today’s DNA synthesis performance is
several orders of magnitude less capable than that for incumbent media (Figure 71). Per the TCO
discussion above, it is probably unnecessary that DNA data write speed be directly competitive
with incumbent media. For example, deep/long archival users may be content with write
throughput in the range of a few 100 megabytes per day (few kilobytes/s). That said, regardless of
the entry point, solutions will need some minimum performance characteristics at a cost that
facilitates a competitive TCO given the use case.

Media Write Latency (time to 1st byte) Write Throughput

DNA Data Storage Minutes to hours ~100 MB/day = 0.001 MB/s
Tape Seconds to minutes ~400 MB/s (uncompressed)
HDD Up to tens of seconds ~300 MB/s
SSD Single digit seconds or less ~500 MB/s (non-NVMe)
Flash Single digit seconds or less ~1000 MB/s
Figure 71. Write latency and throughput for various storage solutions: DNA Data Storage Alliance

Parallelism is foundational to advancing DNA synthesis throughput. Since the speed of chemical
reactions is inherently bounded, a common approach to increasing throughput is to take advantage

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

of the small size of DNA molecules to execute millions to billions of synthesis operations in
parallel.

To spur DNA synthesis innovation for DNA data storage, the IARPA Molecular Information
Storage program (MIST) set synthesis goals (Figure 72) and challenged industry and academia to
develop solutions.

Figure 72. 2022 IARPA Roadmap for DNA synthesis, courtesy David M. Markowitz. Assumes
single stranded DNA, 150 nucleotides in length (20 nt flanking primers), encoded at 1
bit/nucleotide.

BASE-BY-BASE SYNTHESIS
Electrochemical array-based synthesis, in which the fluidics necessary to do synthesis are
combined with CMOS semiconductor technology to control synthesis chemistry on discrete and
very small areas on a chip9,10, is one of the main modalities being pursued. In 2021, Nguyen et al11
showed synthesis of synthetic DNA in an electrochemical array, successfully synthesizing 150
nucleotide long sequences in 650nm wells, at a feature pitch of 2um, or a density of 32 million
synthesis spots per cm2.

Key to the success was demonstrating the ability to localize acid diffusion at each synthesis site,
which is critical to keeping the synthesis chemistry localized to each well and thus ensuring that
the synthesis process is precisely controlled with acceptable error rates. This effort advanced the
state of the art in array-based synthesis density by nearly 3 orders of magnitude. While this chip
was not scaled to full production, if it were, at the densities demonstrated, such a device could
yield performance on the order of a few kilobytes/s/cm2 (assuming each unique DNA sequence
encodes 10 bytes of data and is written over 24 hours) or a few hundreds of megabytes per day.
While far lower than incumbent media, this might be adequate for deep/cold archival data storage
applications. In evidence of further progress, Twist Bioscience stated in 2022 that they had
developed a chip which has a capacity of 1GB per run with about 100M synthesis spots,

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

approaching the trajectory of the MIST goals. Other base-by-base synthesis techniques are also
being pursued, such as ink-jet printing12,13 and light directed synthesis14,15.

Figure 73. Electrochemical DNA synthesis on a nanoscale array. (c) An overview of the nanoscale
DNA synthesis array with scanning electron microscopy images of the 650-nm electrode array and
enlarged view of one electrode. (e) Illustration of the wells patterned with ssDNA oligos with
multiple copies of each oligo per synthesis location.11

DNA ASSEMBLY BASED SYNTHESIS

In addition to the above techniques, which write molecules base-by-base, typically in strands
containing 150-200 nucleotides, there are synthesis techniques being pursued by companies such
as Catalog and Biomemory, where a library of predefined sequences, somewhat akin to movable
type in a mechanical printing press (i.e., symbols) is used to encode the sequences. The motivation
is that these symbols (themselves built with base-by-base synthesis, but needing to be built only
once), in combination with certain encoding strategies, can then be assembled into longer strands
of DNA by various DNA assembly techniques (e.g. ligation) that can enable different synthesis
efficiencies than base-by-base methods.16 Catalog currently claims a peak throughput of about a
terabit per day in their first generation platform (i.e., in the hundreds of megabytes per day range
that was estimated for a scaled version of the electrochemical array-based technology above).

NON-SEQUENCE BASED TECHNIQUES

Lastly, there is research into synthesis techniques which encode data not in the sequence of bases
in the synthetic DNA molecules, but by using (and guiding) the self-assembly properties of DNA
molecules to assemble molecular structures in which the structure itself encodes information; so-
called “Structure-Based DNA Data Storage”17. These techniques are not as far along in scaling for
data storage as are sequence-based techniques, but they represent active areas of work providing
yet further potential alternatives in throughput and information density.

In summary, while significant scaling advances are still required, the fundamentals of synthesizing
DNA encoded with digital data on scalable technology platforms have been shown to work at
speeds which could be practical for early deep/cold archive applications, and substantial
investment in R&D continues within the ecosystem.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

READING DNA (SEQUENCING)

At present, the two primary modes of DNA sequencing most promising for DDS are Sequencing
by Synthesis (SBS), Figure 74, and Nanopore sequencing, Figure 75.

SBS detects DNA bases indirectly. It gets its name because it starts with a single-stranded template
strand of DNA and then synthesizes a complementary DNA strand from that template. As each
base is added in the complementary strand it is identified (typically optically), and the base in the
original template strand can then be identified.

Nanopore sequencing detects bases directly. A strand of DNA is passed through a pore in a
membrane surrounded by an electrolyte solution. With an electrical bias applied across the
membrane, the DNA strand moves through the nanopore, the disruption of ionic current or
tunneling current is registered, enabling the direct detection of the bases in the strand. Nanopores
today are almost entirely biological; however, research is underway to create solid state nanopore
sequencing18.

In general, today, on a raw, per-base basis, SBS is more accurate/slower and nanopore is less
accurate/faster. In terms of scaling in production, there is a battle for throughput/cost. We’ll
examine both.

Figure 74. Sequencing by Figure 75. Nanopore Sequencing:

Synthesis: Illumina National Human Genome Research
Institute (NHGRI)

As tracked by NHGRI (Figure 76) DNA sequencing underwent a rapid cost reduction (and
performance increase) curve starting in 2008, due to the advent of SBS; however, the rate of change
has slowed and leveled out at just under $10/GBase. Illumina stated in 2022 that that their highest
end products can achieve roughly $6/Gbase, that there is a “direct line of site” to $1/Gbase, and
that there are “no conceptual hurdles” to $0.1/Gbase. Assuming we encode 1 bit/base, the
$0.1/Gbase milestone is probably on the horizon over the next 3-5 years and would bring us to a
price of roughly $800/terabyte. This is still quite high for commercial deployment, and more cost
scaling is needed, but note that estimates of DNA sequencing costs today are all based on
medical/scientific use cases and business models, and various factors could further mitigate cost:

• Data storage applications can accept higher error rates than medical/scientific applications.
For example, the cost of sequencing the human genome is frequently used as a benchmark

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

for sequencing cost. In general, the coverage factor (average # of times each nucleotide in
the sequence is sampled to get the final read done) is 30X. In data storage, the DNA codec
can both avoid nucleotide sequences that cause problems during sequencing and can also
recover from errors during decode. Such factors might, for example, enable coverage
factors nearer to 10X. This, in turn, would result in significantly higher effective
throughput during sequencing, with attendant lower costs.

• Related to the previous point, the 1bit/base encoding efficiency used in the estimates here
may be overly conservative; densities between 1 and 2 bits/base would raise effective
throughput and thus lower cost.

• The highest performance form of sequencing today is SBS. While further enhancements
in SBS are inevitable, nanopore sequencing is a promising new technology which,
especially with the advent of solid state nanopore development, could achieve
breakthroughs that initiate another major downward cost (and upward performance) trend.

• Ecosystem considerations and business models could help drive lower sequencing prices
as/if data storage begins driving demand for solutions, providing new demand to the
ongoing and increasing medical/scientific demand.

Figure 76. Cost per Raw Megabase of DNA Sequencing, DNA Sequencing Costs: Data from the
NHGRI Genome Sequencing Program (GSP) Available at: www.genome.gov/sequencingcostsdata

Let’s look at raw performance to get a sense of how much data, and how quickly, we can read
from DNA. The fastest sequencing solutions today are approaching 7 terabases/day which, again
using 1bit/base, is approaching 1 terabyte/day. Is this enough for deep archival recovery? Perhaps.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

Is it enough for more, so-called, “Active Archive” use cases, where reads are more frequent?
Probably not; these use cases will likely require read speeds on order of a few hundred
terabytes/day or higher.

As with DNA synthesis, the costs and performance of DNA sequencing need to continue scaling
to become cost competitive in solutions; however, we are likely approaching the requirements of
some use cases, and, also as with synthesis, the fundamental technologies have been demonstrated,
and scaling is now the task.

SUSTAINABILITY
Another important attribute of DDS is sustainability. At first order, DDS has inherent advantages
for sustainability due to its extremely high data density and small volumetric footprint. In
comparison to sprawling data centers, there is simply not much of a comparison.

More systematically, the most important metrics that contribute to sustainability are (1) greenhouse
gas emissions (kgCO2equivalents/TB), (2) energy consumption (MJ/TB), and (3) water
consumption (L/TB). Recent life cycle assessment (LCA) simulations19,20 have indicated that the
sustainability profile of DDS approaches compare well to tape and other incumbent storage media.
This will have to be demonstrated at scale.

The major sustainability issue for DDS today is the use of oil-based chemicals needed for
Phosphoramidite synthesis, which are toxic and flammable; however, the projection is that
enzymatic synthesis, which uses aqueous based processes, will mitigate this issue.

BIOSECURITY
It is important to note that no organisms or living cells are required for DNA data storage; synthetic
DNA for data storage is constructed and manipulated through well-controlled chemical processes,
and the sequences generated for it are controlled by software codecs. It is already the case that
organizations, such as the International Gene Synthesis Consortium21, have been formed to both
conform with governmental regulations, and further enhance sequence screening protocols to
reduce the risk of generating sequences associated with pathogenic organisms. Moreover, due to
the fact the sequences encoded for data storage are not constrained by biological/medical
requirements, DNA codecs can and will embed biosecurity requirements as they exist today, and
as required with the evolution of the DNA data storage ecosystem.

SUMMARY
Synthetic DNA as a storage medium is compelling due to its ~1bit/nm3 bit size, its molecular
stability, and the property that it is decoupled from the read/write device. These aspects, if
delivered in commercial systems, can enable compelling TCO for high capacity and long time
scale archival use cases, as well as potentially for other use cases as the capabilities of the
technology evolve.

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

reasonable to expect that a path to a commercial DNA data storage ecosystem will come into focus
over the next 5-10 years.

REFERENCES

1
Gartner Market Trends: Evolving Enterprise Data Requirements —How Much Is Not Enough? (ID G00727554)
2
Arcilla, A. Quantifying the Economic Benefits of LTO-8 Technology.
3
TCO Calculator for Data Storage | Fujifilm [United States]. https://www.fujifilm.com/us/en/business/data-
storage/resources/tco-tool.
4
Preserving our digital legacy: An introduction to DNA data storage”, DNA Data Storage Alliance;
https://dnastoragealliance.org/dev/wp-content/uploads/2021/06/DNA-Data-Storage-Alliance-An-Introduction-
to-DNA-Data-Storage.pdf
5
Organick et al. An Empirical Comparison of Preservation Methods for Synthetic DNA Data Storage. Small
Methods 5, 2001094 (2021).
6
Grass RN, Heckel R, Puddu M, Paunescu D, Stark WJ. Robust Chemical Preservation of Digital Information on
DNA in Silica with Error-Correcting Codes. Angew Chem Int Ed Engl. 2015; 54(8): 2552-2555
7
Coudy, D., Colotte, M., Luis, A., Tuffet, S. & Bonnet, J. Long term conservation of DNA at ambient
temperature. Implications for DNA data storage. PLOS ONE 16, e0259868 (2021).
8
Bonnet, J. et al. Chain and conformation stability of solid-state DNA: implications for room temperature
storage. Nucleic Acids Res. 38, 1531–1546 (2010).
9
Kosuri, Sriram, and George M. Church. "Large-scale de novo DNA synthesis: technologies and applications."
Nature methods 11.5 (2014): 499-507
10
Jensen M.A., Davis R.W, Template-Independent Enzymatic Oligonucleotide Synthesis (TiEOS): Its History,
Prospects, and Challenges; Biochemistry 2018, 57, 12, 1821–1832, March 13, 2018;
https://doi.org/10.1021/acs.biochem.7b00937
11
Nguyen, B. H. et al. Scaling DNA data storage with nanoscale electrode wells. Sci. Adv. 7, eabi6714.
12
Catalog, https://www.catalogdna.com
13
Oligonucleotide Library Synthesis | Agilent. https://www.agilent.com/en/product/sureprint-oligonucleotide-
library-synthesis/oligonucleotide-library-synthesis
14
Efcavitch, J. W. & Holden, M. T. Homopolymer encoded nucleic acid memory (2021); US Patent #: US-
11174512-B2
15
Antkowiak, P.L., Lietard, J., Darestani, M.Z. et al. Low cost DNA data storage using photolithographic
synthesis and advanced information reconstruction and error correction. Nat Commun 11, 5345 (2020).
https://doi.org/10.1038/s41467-020-19148-3
16
Roquet et al. DNA-based data storage via combinatorial assembly. https://doi.org/10.1101/2021.04.20.440194
17
Doricchi et al. Emerging Approaches to DNA Data Storage: Challenges and Prospects. ACS Nano 2022 16
(11), 17552-17571 DOI: 10.1021/acsnano.2c06748
18
Xue, L., Yamazaki, H., Ren, R. et al. Solid-state nanopore sensors. Nat Rev Mater 5, 931–951 (2020).
https://doi.org/10.1038/s41578-020-0229-6
19
Nguyen, B. H. et al. Architecting Datacenters for Sustainability: Greener Data Storage using Synthetic DNA.
20
Mytton, D. Data centre water consumption. Npj Clean Water 4, 1–6 (2021).
21
https://genesynthesisconsortium.org

THE INTERNATIONAL ROADMAP FOR DEVICES AND SYSTEMS: 2023

RoadMap To Semiconductor Tech 1997
No ratings yet
RoadMap To Semiconductor Tech 1997
239 pages
MODERN ENTERPRISE Data Pipeline
100% (1)
MODERN ENTERPRISE Data Pipeline
98 pages
Storage 3.0: Storage Innovation Enters An Exciting New Phase
No ratings yet
Storage 3.0: Storage Innovation Enters An Exciting New Phase
171 pages
h19267 WP Idc Storage Reqs Digital Enterprise
No ratings yet
h19267 WP Idc Storage Reqs Digital Enterprise
17 pages
HP Simply Storageworks: Introduction To Storage Technologies
100% (2)
HP Simply Storageworks: Introduction To Storage Technologies
16 pages
2022irds Es
No ratings yet
2022irds Es
86 pages
2022IRDS BeyondCMOS
No ratings yet
2022IRDS BeyondCMOS
132 pages
2021irds BC
No ratings yet
2021irds BC
129 pages
Data Center Vision How Data Center Infrastructure Will Evolve To Support Ai and Accelerated Compute Us53192425 Ib
No ratings yet
Data Center Vision How Data Center Infrastructure Will Evolve To Support Ai and Accelerated Compute Us53192425 Ib
25 pages
2020irds BC
No ratings yet
2020irds BC
127 pages
2021irds Es
No ratings yet
2021irds Es
64 pages
2023irds Fac
No ratings yet
2023irds Fac
77 pages
2022IRDS MoreMoore
No ratings yet
2022IRDS MoreMoore
39 pages
2021IRDS MTM
No ratings yet
2021IRDS MTM
47 pages
2022IRDS WP-MorethanMoore
No ratings yet
2022IRDS WP-MorethanMoore
62 pages
2009 Links and Table List
No ratings yet
2009 Links and Table List
47 pages
2023irds BC
No ratings yet
2023irds BC
131 pages
TOFC
No ratings yet
TOFC
17 pages
I R D S ™: Nternational Oadmap FOR Evices and Ystems
No ratings yet
I R D S ™: Nternational Oadmap FOR Evices and Ystems
47 pages
Assignment 2
No ratings yet
Assignment 2
17 pages
Digital Storage Memory Technology PDF
No ratings yet
Digital Storage Memory Technology PDF
31 pages
Current Trends in Computer Hardware & Software Platfroms
No ratings yet
Current Trends in Computer Hardware & Software Platfroms
11 pages
Storage Concepts Storing and Managing Digital Data
No ratings yet
Storage Concepts Storing and Managing Digital Data
344 pages
The Paan
No ratings yet
The Paan
14 pages
ToC - Advanced Materials For Electronics Market - Global Industry A...
No ratings yet
ToC - Advanced Materials For Electronics Market - Global Industry A...
12 pages
De Jesus CSPT C9
No ratings yet
De Jesus CSPT C9
7 pages
I R D S: Nternational Oadmap FOR Evices and Ystems
No ratings yet
I R D S: Nternational Oadmap FOR Evices and Ystems
23 pages
CPH Micro
No ratings yet
CPH Micro
14 pages
297 - Gartner 3Q23 Semiconductor Forecast Webinar-Version Final 20231012
No ratings yet
297 - Gartner 3Q23 Semiconductor Forecast Webinar-Version Final 20231012
25 pages
SRC Mapt Roadmap 2023 v4
No ratings yet
SRC Mapt Roadmap 2023 v4
212 pages
Untitled Document
No ratings yet
Untitled Document
9 pages
Itc Ralph
No ratings yet
Itc Ralph
4 pages
SEMICON - Memory Technology and Overall Trends in The Semiconductor Industry
No ratings yet
SEMICON - Memory Technology and Overall Trends in The Semiconductor Industry
14 pages
Battery Management System Market: Global Opportunity Analysis and Industry Forecast, 2025-2034
No ratings yet
Battery Management System Market: Global Opportunity Analysis and Industry Forecast, 2025-2034
49 pages
International Roadmap For Devices and Systems 2022 Lithography (IEEE)
No ratings yet
International Roadmap For Devices and Systems 2022 Lithography (IEEE)
15 pages
Ieee CS: Read On
No ratings yet
Ieee CS: Read On
163 pages
PWC - Semiconductor-Opportunities For Market - Global-2019.4 PDF
No ratings yet
PWC - Semiconductor-Opportunities For Market - Global-2019.4 PDF
18 pages
5 - 2015 ITRS 2.0 - More Moore
No ratings yet
5 - 2015 ITRS 2.0 - More Moore
52 pages
State of The Semiconductor Industry Report
No ratings yet
State of The Semiconductor Industry Report
33 pages
CMP 112 Assignment
No ratings yet
CMP 112 Assignment
7 pages
Report Computer
No ratings yet
Report Computer
9 pages
Emerging Trends and The Outlook For Data Storage in The Future
No ratings yet
Emerging Trends and The Outlook For Data Storage in The Future
16 pages
De Unit-4
No ratings yet
De Unit-4
20 pages
Microelectronics PDF
No ratings yet
Microelectronics PDF
15 pages
2020irds MM
No ratings yet
2020irds MM
37 pages
PC Industry
No ratings yet
PC Industry
2 pages
10 Tech Industry Trends For 2023
No ratings yet
10 Tech Industry Trends For 2023
10 pages
Ilovepdf Merged
No ratings yet
Ilovepdf Merged
50 pages
Trends in Technology
No ratings yet
Trends in Technology
2 pages
(Backup) 2025 Preview Market Intelligence Report - Active ICs ©2025 N.F. Smith & Associates, L.P.pdf..20250313234629818
No ratings yet
(Backup) 2025 Preview Market Intelligence Report - Active ICs ©2025 N.F. Smith & Associates, L.P.pdf..20250313234629818
26 pages
Computer Market
No ratings yet
Computer Market
2 pages
Assignment Serie1
No ratings yet
Assignment Serie1
7 pages
Technology Prospects For Data-Intensive Computing
No ratings yet
Technology Prospects For Data-Intensive Computing
21 pages
Graph Santiago 2
No ratings yet
Graph Santiago 2
2 pages
Type of Storage - 075322
No ratings yet
Type of Storage - 075322
9 pages
Semiconductors Sector Report
No ratings yet
Semiconductors Sector Report
2 pages
List of Top 100 BSE and NSE Listed Companies Arranged by Sectors
No ratings yet
List of Top 100 BSE and NSE Listed Companies Arranged by Sectors
73 pages
Roadmap Semiconductor Manufacturing
No ratings yet
Roadmap Semiconductor Manufacturing
16 pages
RCDD Pre-Examination Assessment
100% (1)
RCDD Pre-Examination Assessment
19 pages
Touchless Touch Screen Report
67% (3)
Touchless Touch Screen Report
24 pages
JIG HSSEMS Standard Issue
No ratings yet
JIG HSSEMS Standard Issue
23 pages
957 4 1 A 57607 ECAD VLSI Lab Manual PDF
No ratings yet
957 4 1 A 57607 ECAD VLSI Lab Manual PDF
151 pages
Basic Networking Hardware: Kamal Harmoni Kamal Ariff
100% (2)
Basic Networking Hardware: Kamal Harmoni Kamal Ariff
22 pages
DAY - 1 Day - 6 DAY - 11: Extended Primitives Modelling
No ratings yet
DAY - 1 Day - 6 DAY - 11: Extended Primitives Modelling
2 pages
Poker Math
No ratings yet
Poker Math
212 pages
Virtual Organization
No ratings yet
Virtual Organization
11 pages
Nokia L3 Data Collector For LTE (MegaPlexer) User Manual v18.2.2
No ratings yet
Nokia L3 Data Collector For LTE (MegaPlexer) User Manual v18.2.2
76 pages
Three Phase Inverter With Synergy Technology: SE50K / SE55K / SE82.8K
No ratings yet
Three Phase Inverter With Synergy Technology: SE50K / SE55K / SE82.8K
2 pages
Dissertation
No ratings yet
Dissertation
62 pages
AI Prompts 1
No ratings yet
AI Prompts 1
5 pages
Gs36j02a10-01e 047
No ratings yet
Gs36j02a10-01e 047
13 pages
Gujarat Technological University: Page 1 of 3
No ratings yet
Gujarat Technological University: Page 1 of 3
3 pages
Teradata Vantage SQL Basics
No ratings yet
Teradata Vantage SQL Basics
14 pages
Cheriyamundam Grama Panchayat: Transaction Details
No ratings yet
Cheriyamundam Grama Panchayat: Transaction Details
1 page
Bcse209l Machine-Learning TH 1.0 0 Bcse209l
No ratings yet
Bcse209l Machine-Learning TH 1.0 0 Bcse209l
3 pages
All About Wind Synthesizers
100% (1)
All About Wind Synthesizers
5 pages
A Survey On Generative Diffusion Model
No ratings yet
A Survey On Generative Diffusion Model
25 pages
Hardware Manual of T20
No ratings yet
Hardware Manual of T20
10 pages
The Inmate of The Dungeon1894 by Morrow, W. C.
No ratings yet
The Inmate of The Dungeon1894 by Morrow, W. C.
16 pages
M2L2 Producing Reports Exercises
No ratings yet
M2L2 Producing Reports Exercises
6 pages
Employees Registration in HH - User Manual
No ratings yet
Employees Registration in HH - User Manual
8 pages
1) Functional Units of Computer
No ratings yet
1) Functional Units of Computer
6 pages
ACR Trafo
No ratings yet
ACR Trafo
14 pages
Implement Basic Connectivity
No ratings yet
Implement Basic Connectivity
9 pages
AN1002 Getting Started With HCS12
No ratings yet
AN1002 Getting Started With HCS12
21 pages
Ethical Hacking
No ratings yet
Ethical Hacking
2 pages
Optical Isolation Amplifiers Support Inverter Voltage, Current and Temperature Sensing
No ratings yet
Optical Isolation Amplifiers Support Inverter Voltage, Current and Temperature Sensing
6 pages
Cubic Spline v101
No ratings yet
Cubic Spline v101
2 pages
Securing ChatGPT: Best Practices for Protecting Sensitive Data in AI Language Models
From Everand
Securing ChatGPT: Best Practices for Protecting Sensitive Data in AI Language Models
Matthew C. Smith
No ratings yet
Software Patterns Made Easy
From Everand
Software Patterns Made Easy
Justice Nanhou
No ratings yet