Lec21 PDF

The University of Texas at Dallas
Erik Jonsson School of Engineering and

Computer Science
Memory Management in Modern Computers

One of the biggest problems facing modern computer designers is
that of providing large amounts of high speed memory.
This is a problem that has evolved over the last 25-30 years.
Earlier in the history of computing, most processors were
relatively slow compared to the speed of available memories
(Except for bulk storage mechanical memories, i.e., disks and
drums).
Especially in the early days of the personal computer, the CPU
was relatively slow compared to early electronic memories.
The speed of random-access memory was not an issue; the biggest
problem was just getting enough memory, period (early PCs with
large memories had 256-512 Kbytes)!
1
Lecture # 21: Memory Management in Modern Computers
N. B. Dodge 9/15

Computer Science
Relative Speeds of the CPU and DRAM

Over the last two+ decades, central processor chips have caught up
with and passed DRAM speed dramatically.
Example: current CPU speed is 3-4.5 GHz, depending on the
processor type, and should increase somewhat, although
manufacturers are now abandoning the speed race in favor of
multiple processors.
On the other hand, practical bus speed for CPU memory is about 11.8 GHz currently, and this is for high performance memory;
common bus speeds are still no more than half that.
The CPU performance edge over memory is on the order of 3-4, and
much more than that on systems with the more common bus speeds.
2
N. B. Dodge 9/15

Computer Science
The Memory Speed/Cost Dilemma

There are further problem facing the modern computer
designer:
Users need very high memory speeds to improve performance
(for example, in graphical computing, games, video editing).
At the same time there is also great demand for maximum
memory capacity by many users (PCs do not just manipulate text
any more; complex graphics, video games, movie editing and
animation all require enormous amounts of both DRAM and bulk
storage (hard drives [HDDs]).
However, there is a conflict in these requirements:

Fast memories are very expensive.
High-capacity, cheap memories (esp. HDDs) are very slow.
3
N. B. Dodge 9/15

Computer Science
Stating the Memory Management Problem

The computer designer of today is therefore faced with
a problem that is not easy to solve:
There must be enough high-speed memory available to avoid
slowing down the processing rate of current CPUs.
There must be sufficient DRAM to avoid the deadly disk
access (i.e., having to go to the HDD to get program or data
material), at least very often, since HDD access is very slow.
There must be enough bulk memory (HDD) for all storage
needs, and accessing this memory and transferring it to
DRAM/other memory must be as painless as possible.
The cost must be reasonable.
4
N. B. Dodge 9/15

Computer Science
Solving the Problem of Memory Management

The current approach to memory management :
CPU has a large register complement, which allows more data in the
CPU at a time and improves performance.
Very-high-speed D flip-flop arrays, called cache, hold currently
executing program segments. There are two kinds:
L1/L2 cache On CPU chip, adjacent to ALU. ~ 64 Kbytes, very fast.
L3 cache Center of CPU chip.
~ 15 Mbyte, very fast.
High-speed electronic memory (DRAM, up to 64 Gbytes, fast)

provides capacity for programs currently in process.
Bulk storage memory (disk drives, ~0.3-2+ TByte, slow but cheap)
holds complete programs and near-term archives.
Slower memories such as CDs and DVDs for long-term storage.
5
N. B. Dodge 9/15

Computer Science
Types of Memory
As we have just seen, even in the everyday PC, the use of
sophisticated memory management is common.
This means that there are four kinds of memory in the modern PC
or workstation computer: Registers, cache, DRAM, and the disk
or HDD (or SSM). And this does not count CDs, DVDs, Zip
drives, thumb drives (flash EPROM), or floppy disks!
The challenge to the computer engineer is to mesh the first five
storage media and to make the use of them transparent that is,
invisible to the user, who will appear to have massive amounts of
high-speed, cheap memory available to solve any problem.
Before we discuss how to manage this extremely challenging
engineering problem, we will discuss the types of memory that are
used and learn a little about them.
6
N. B. Dodge 9/15

Computer Science
Registers
D Q
CR
D FF
32-Bit Reg.
Register Block
We already know that registers are simply collections of D FFs.

Most CPUs today contain many registers, (e.g. the R-2000s 32).
Registers are inside the CPU, adjacent to the ALU, so their speed
is basically that of the CPU (in fact, they determine ALU speed).
7
N. B. Dodge 9/15

Computer Science
Random-Access Electronic Memory

Random-access memories (RAM) make up the working memory
of most computers.
These memories are referred to as random-access because the
entire array of memory is immediately available to be used; any
single byte in the memory may be loaded or stored (randomly
accessed) in the same amount of time.
There are two primary types of RAM: Static RAM (SRAM), and
dynamic RAM (DRAM). Both SRAM and DRAM are used in
modern computers such as the PC.
SRAM is used in what are referred to as caches small, very-highspeed memories that are physically close to the CPU.
DRAM, though very fast, is slower than SRAM, but because it is
inexpensive, it is the primary memory in most personal computing
systems.
8
N. B. Dodge 9/15

Computer Science
L1 Cache
L1 cache (level 1 cache) is SRAM memory that is very close to
the CPU. For example, it is next to the ALU in most processors.
L1 cache is basically sets of D FFs but many more than in the
CPU register block.
For example, a typical register block might have 16-32 registers of
4 or 8 bytes each for a total of 64-128 bytes of storage. The Intel L1
Core i7 cache, on the other hand, has 64 Kbytes (32K instruction,
32K data) the equivalent of ~ 500,000 flip-flops.
Access speed of L1 cache is slower, however, due to the complex
arrangement of data buses which is necessary to access specific
bytes in the L1 memory array. It is typically about 1/2-1/3 as fast
as CPU registers in terms of load/store cycle.
9
N. B. Dodge 9/15

Computer Science
L1 Cache (Continued)
D Q
CR
D FF
32-Bit Reg.
L1 Cache
As you saw on the previous slide, in terms of memory structure,

cache has regressed. Modern computer chips have separate
instruction and data caches!
10
N. B. Dodge 9/15

Computer Science
L2 Cache
The level-2 cache is a bit farther away on the
chip. (L2 is also SRAM).
L2 cache is much larger, since more real
estate is devoted to memory. The Intel Core
i7 has 1 Mbyte cache.
Due to even more elaborate bus arrangements
and the fact that L2 cache is not as close to the
CPU, load/store access is > L1 cache.
11
N. B. Dodge 9/15

Computer Science
L3 Cache
In modern multicore processors, the cores share the L3
cache, which is typically 8-12 Mbyte.
As L1 cache is slower than the register block, and L2 is
slower than L1, L3 cache is slower still, though much
faster than DRAM. The reason is that the L3 cache is
yet even farther away from the CPU, though still on the
chip.
To minimize the degradation in memory speed, the
CPUs are typically clustered around the L3 cache, as
shown in the picture of the Core i7 chip (upcoming).
12
N. B. Dodge 9/15

Computer Science
Intel Core I7 Cache Structure
13
N. B. Dodge 9/15

Computer Science
Processor Layout of Single Intel Core i7 CPU
14
N. B. Dodge 9/15

Computer Science
Cache Location on Intel CPU
15
N. B. Dodge 9/15

Computer Science
Why Not More Cache?

The question arises: If cache memory is so great, why
isnt all computer memory fast cache?
Answer: Cache memory has two major problems:
It consumes huge amounts of power compared to DRAM
memory (a flip-flop has about sixteen transistors; a DRAM cell
uses only one).
This means if more cache were used, the cost of a computer
(think PC) would go up dramatically, due to the cost of extra
power to run it, and cost of cooling the computer!
Also, cache is much more expensive than DRAM (5:1 or more).
For that reason, DRAM memory is an excellent

compromise solution to fast storage problems.
16
N. B. Dodge 9/15

Computer Science
Comparison of SRAM and DRAM

SRAM
Parameter
DRAM
Very fast
High; ~16
transistors
per storage cell
High
Speed
Fast
Low; 1
transistor
per cell
Very low
Virtually
none
Very low
Excessive
High
17
Complexity
Power Used
Heat
Generated
Cost
N. B. Dodge 9/15

Computer Science
DRAM Memory
The term DRAM stands for dynamic random-access memory
(pronounced D-ram, not dram). This means that the title
above is actually redundant!
DRAM is electronic memory that is capable of very fast access
(load or store), but is not as fast as cache. One exception is
Rambus memory, a special DRAM memory whose
manufacturer has announced cache-speed products (up to 7.2
GHz!). It is very expensive, however.
The simple construction of DRAM makes it ideal in modern,
workstation-based computing, where most users have their own
computer system (PC, Mac, Sun, etc.).
DRAM consists of a simple charge-storage device (stored charge =
1), with a switch to store/test the charge. Only a single
transistor is required for a DRAM bit cell.
18
N. B. Dodge 9/15

Computer Science
DRAM (Continued)
The term dynamic in DRAM is due to the fact that the
memory is not truly a flip-flop; it is not static. DRAM
remembers a 1 by storing charge on a capacitor.
Capacitors, however, are not perfect storage elements
the charge leaks off after a short time. Thus the DRAM
element is dynamic its memory lifetime is limited
and it must have its memory refreshed periodically.
On the next several slides, we explore the way DRAM is
constructed and the odd way that it must be treated to
be sure that it retains its memory.
19
N. B. Dodge 9/15

Computer Science
DRAM Memory Cell Construction

Bit line
CMOS
transistor
Word
line
Capacitor
Ground
The DRAM cell is quite simple, consisting of a single CMOS

transistor and a capacitor, which can store electronic charge.
The capacitor is grounded on one end. Wires connect two
terminals of the transistor to lines that can apply voltage.
20
N. B. Dodge 9/15

Computer Science
DRAM Cell Operation

Bit line +V (= logic 1)
Bit line
Current
+V (= logic 1)
turns on
transistor
Word
line
CMOS
transistor
Capacitor
charges
Current
+V (= logic 1)
turns on
transistor
Word
line
CMOS
transistor
Ground
To write logic 1 data to a DRAM cell, a

voltage is applied to the word line, which turns
the transistor on (it is like an electronic
switch). If a voltage V is applied to the bit
line, current flows into the capacitor and
charges it, creating a logic 1.
21
0V (= logic 0)
Capacitor
discharges
Ground
To write logic 0 data to a DRAM cell, a

voltage is applied to the word line, which turns
the transistor on (once again, like an electronic
switch). Now, if 0 volts (ground) is applied to
the bit line, current flows out of the capacitor
and discharges it, creating a logic 0.
N. B. Dodge 9/15

Computer Science
DRAM Cell Operation (2)

Bit line logic 1 sensed
Bit line
No current flow
Current
+V (= logic 1)
turns on
transistor
Word
line
CMOS
transistor
Capacitor
charged
+V (= logic 1)
turns on
transistor
CMOS
transistor
Word
line
Ground
Read 1 memory cycle.
logic 0 sensed
Capacitor
has no charge
Ground
Read 0 memory cycle.
To read, or sense the value of the DRAM cell, the word line once again has a
voltage applied to it, which turns on the transistor. If the capacitor is charged,
current flows OUT of the transistor, and this current is sensed and amplified,
showing that a 1 is present. If the capacitor is discharged, no current flows,
so that the sensing element determines that a logic 0 is present.
22
N. B. Dodge 9/15

Computer Science

Note that in reading a DRAM memory cell with a 1 in it (charge
stored on capacitor), the act of reading destroys the 1 by
draining the charge off the capacitor.
Therefore, after reading a 1, it must be rewritten.
Also, as time passes, whether used or not, the capacitor loses
charge so that the logic 1 eventually disappears.
We see that even if a 1 is not read, the charge must be periodically
replaced or the DRAM memory loses its mind!
In a modern DRAM cell, this refresh must occur every few
milliseconds.
The refresh cycle is not long, however, taking 4-5% of total
memory read/write time, which does not reduce memory speed or
efficiency to any great degree.
23
N. B. Dodge 9/15

Computer Science

Bit line
Logic 1 read (or sensed in

refresh cycle) by draining capacitor
Current
Word line
activated
Word
line
CMOS
transistor
+0
Capacitor
discharges
Bit line
Current
Word line
reactivated
CMOS
transistor
Word
line
0+
Capacitor
recharged
Ground
Read 1 memory cycle or refresh

cycle logic 1 detect.
Logic 1 rewritten by
applying +V to bit line
Ground
Read or refresh cycle

logic 1 rewrite.
The refresh cycle occurs after a logic 1 read or periodically if the

memory cell is not accessed. The refresh cycle is typically every
few milliseconds. Obviously if the cell is a 0, it is not recharged.
24
N. B. Dodge 9/15

Computer Science
DRAMs Get Denser
Current new DRAMs top the

scale at 64 and 128 gigabits!
25
N. B. Dodge 9/15

Computer Science
Exercise 1
1. Rank these memories by speed: L2 cache, DRAM, L1
cache, registers, and hard disk drives.
2. A DRAM memory chip is accessed and a bit read out.
The bit that is read is a 1. What happens now?
3. That same memory bit is then left alone (i.e., not
accessed by its addressing mechanism for either read
or write) for several milliseconds. What happens
next?
26
N. B. Dodge 9/15

Computer Science
Bulk Storage (Disk Storage or HDD)

Electromechanical data storage is normally not random-access like
SRAM or DRAM.
This means that data cannot normally be accessed in arbitrary
order, but must be loaded or stored according to rules, which
generally have to do with positioning a recording mechanism over
the correct location in an expanse of recording media prior to
being able to perform the memory access.
That is, the correct segment of data must be located (normally by
mechanically moving a recording head) before it can be read.
This load/store operation is particularly time-consuming, because
it involves mechanical movement rather than simply electronic
switching.
28
N. B. Dodge 9/15

Computer Science
HDD Read/Write Mechanism

Strong, concentrated
magnetic field
Rotation
Aluminum disk coated

with magnetic material
Magnetic field lines

(direction depends
on current flow)
Current flow
The HDD stores data on a rotating disk coated with magnetic

material.
A magnetic coil is used to record each one and zero. Current in the
coil generates a magnetic field, which magnetizes material in the
HDD surface. One direction of current writes a 1, the other a 0.
When the coil is later positioned over the disk to read, the oppositepolarity 1s and 0s cause back-and-forth current flow according to
whether a 1 or 0 is present. In this way, the data is detected.
29
N. B. Dodge 9/15

Computer Science
Hard Disk Drive Example

Portion of
read/write
electronic
circuitry
(the rest is
on the back
side of the
unit on a
separate
circuit
board).
30
Metal disk covered

with magnetic
coating
Recording head
N. B. Dodge 9/15

Computer Science
Detail of Disk Read/Write Head

Flexible cable
carries signals
to amplifier
circuitry to be
converted to
digital signals
Positioning
mechanism
Positioning arm
Recording head
31
N. B. Dodge 9/15

Computer Science
HDD Side View, Showing Multiple Disk Platters

Upper recording
and reading head;
note that
positioning
mechanism moves
both heads
simultaneously.
Second
recording disk
surface
(recording head
not visible)
32
N. B. Dodge 9/15

Computer Science
HDD Package
The HDD is usually packaged in a metal case. Higher-quality

units are typically packaged in an aluminum casting, or
similar rigid container, which provides stability and better
data integrity.
33
N. B. Dodge 9/15

Computer Science
HDD Storage/Retrieval is Slow

The latency (time to get/store data) of a HDD is given by
the formula:
latency = seek time + rotational delay + transfer time + controller delay
Where:
Seek time = time for the positioning arm to move the head from its
present track to the track where the load/store data is located.
Rotational time = time for the requested sector to rotate underneath
the read/write head after the head is positioned over the track.
Transfer time = time for data transfer from disk to main memory.
Controller delay = time to set up transfer in the HDD electronic
interface.
34
N. B. Dodge 9/15

Computer Science
HDD Storage/Retrieval is Slow (2)

Example: latency of writing one 512-byte sector on a
magnetic disk rotating at 7200 rpm, with the following
parameters:
Average seek time = 12 ms (typical for movement across half
the disk)
Transfer rate = 5 Mbytes/sec; transfer time =
[0.000512Mbyte/5 Mbyte/sec] = 0.1 ms
Controller delay = 2 ms
Rotational time depends on the position of the first byte to be
transferred, but on average will be ([1/7200)]60[1/2]) = 4.2
ms (average rotation = of circle).
Then average latency = 12 ms + 4.2ms + 0.1 + 2 ms =

18.3 ms. Note that actual transfer time is small!
35
N. B. Dodge 9/15

Computer Science
Other Disk Storage Units and Media
Other storage media include CDs, DVDs and thumb drives.

Most of these storage units are removable media.
Floppy disks, hard drives, Zip drives, and tapes are magnetic media.
The CD-ROM and the DVD use optical recording/reading involving
a laser beam to record and read data. They are relatively slow.
The thumb drive uses electronic memory called EPROM
(erasable, programmable read-only memory), and is a true solidstate memory. These drives were the first SSDs.
Very fast EPROMs (SSDs) are beginning to be available for fast
bulk storage, replacing HDDs on laptops. They are relatively
expensive.
36
N. B. Dodge 9/15

Computer Science
SSDs Get Larger and Faster
37
Last year, Samsung launched

the SSD 850 EVO with storage
up to 1 TB and the reviews have
been very positive.
The best budget option is
the SanDisk Ultra II, at 960GB,
with lower performance and
warranty.
Samsung also introduced a 1 TB
850 Pro, with a slightly faster
read speed and best-in-class 10
year
N. B. Dodge 9/15

Computer Science
Biggest and Best SSDs

The biggest solid state drives available in 2015:
1. LSI: 4 TB SSD ($29,000).
2. OCZ Chiron: 4 TB SSD (may be expandable past 4 TB), also
$15-30K.
3. Samsung 850 Pro ($600-650): 1 TB SSD (10-year warranty)
4. Samsung 850 EVO SSD ($400-450): 1 TB SSD (5-year
warranty) Priced for the home user!
5. Crucial M550 ($400-450): 1 TB SSD (3-year warranty)
6. SanDisk Ultra II ($350-400): 960 GB SSD (3 yr. warranty a
decent budget option)
38
N. B. Dodge 9/15

Computer Science
The Memory Hierarchy

We have described a number of memory devices which
are useful for storing and reading computer data.
All of these (other than archival types) are used in a
mix on the modern computer for real-time storage and
retrieval of data.
Since SRAMs the best data storage media if not so
power-hungry and costly cannot be used exclusively,
a mix of L1 and L2 cache, DRAM, and HDDs make up
the memory hierarchy of most computers.
The trick is to design a mix of these types which will
give the highest performance for a reasonable price.
39
N. B. Dodge 9/15

Computer Science
Arrangement of the Memory Hierarchy

Memory arrangements make use of the fact that
programs exhibit two common behaviors:
Temporal locality Recently-used code and data is often
reused (e.g., a loop program continues to use the same steps).
Spatial locality Recently-accessed data items are usually close
to other recently-accessed (or about-to-be-accessed) data items.
Modern schemes use a shuffling methodology that

moves data from slower storage media to faster media.
Higher-speed memories are also placed closer to the
CPU, since memory access also depends on the
proximity of the storage element; electronic signals
propagate at about 33 ps/cm.
40
N. B. Dodge 9/15

Computer Science
Arrangement of Levels in Memory Hierarchy

CPU Package
CPU
L1
Cache
L2/3
Cache
DRAM
HDD
Registers
Size: <300 Bytes

Speed: 100 ps
8-64 Kbytes
200 ps
0.5-3 Mbytes 0.5-4 Gbytes

0.2-0.5 ns
1-10 ns
160-2000 Gbytes
~10-20 ms
Memory is physically arranged so that fastest elements (registers)

are closest to the CPU and slower elements are progressively
farther away.
41
N. B. Dodge 9/15

Computer Science
The Importance of Cache

As mentioned previously, the key to modern computer performance
is not the CPU CPU performance has far-outstripped the speed of
most computer memories.
The key is the use of cache. The secret of todays high-performance
PCs and workstations is the design of an architecture that allows
maximum use of DRAM and HDD (cheap) plus just enough SRAM
cache (expensive and power-consuming), thus enabling the CPU to
realize most of its performance advantage.
The method used is the shuffling technique alluded to two slides
back. This method uses a very high speed, complex arrangement to
constantly move program and data content from slower to faster
memory as the CPU executes a process.
42
N. B. Dodge 9/15

Computer Science
Cache Utilization
Cache designers make use of the principles of temporal and spatial
locality to assure that the most-probably needed instructions and
data are available to the computer in cache (to speed execution).
Special hardware is designed to manage cache content with the
goal of forecasting upcoming instructions and data required by the
processor during program execution and moving it from slower
DRAM into cache.
This hardware has two special goals: (1) examining the currentlyexecuting process and predicting instruction and data need, and
(2) moving the required information from DRAM to cache in a
timely manner to foresee that anticipated need.
43
N. B. Dodge 9/15

Computer Science
Looking for Data/Instructions in the Cache
44
Clearly the purpose of cache management is to make sure that ALL

upcoming instructions and data are in the cache.
This brings up two questions: (1) how does the processor know that data
is in the cache, and (2) if it is NOT there, how does the processor get it and
what sort of performance penalty is there?
There are several ways in which the cache can be assigned DRAM
memory correspondence. The simplest is direct mapping, in which each
block of memory in cache is assigned to some number of DRAM locations.
When a program needs a particular DRAM location to be loaded, it goes
to the corresponding cache location to get the data. This leads to further
complications, in that now we need validity indicators for each cache
location. This is because since each cache block is assigned to several
memory blocks in DRAM, the program needs to know if the right data is
available in cache at the time it is needed.
N. B. Dodge 9/15

Computer Science
Looking for Data (2)

If the correct data is not in cache, the hardware memory manager
declares a cache miss. This means that the program must be
delayed for several clock cycles while the required instruction or
data is moved from DRAM to cache.
We see that a cache miss is highly undesirable, since it can
substantially slow down the program.
A key part of cache memory management, then, is to minimize the
cache misses, which correspondingly increases the speed of
execution of a program.
There are a number of clever and effective cache management
designs, which dramatically reduce cache misses and improve
computer performance. They are, however, beyond the scope of
EE 2310.
45
N. B. Dodge 9/15

Computer Science
Cache Diagram
CPU Package
CPU
L1
Cache
L2
Cache
DRAM
HDD
Registers
Cache management hardware

includes subsystems to predict
usage and move data or
instructions from DRAM to
cache as appropriate. A cache
miss will initiate DRAM access
for transfer to cache.
46
N. B. Dodge 9/15

Computer Science
Summary
Modern memory management maximizes the speed of
computer processing while keeping system cost
reasonable for the user.
This approach uses a small amount of very fast SRAM
cache memory which are physically near the computer,
a substantial amount of DRAM, which is still very fast,
as the main working memory, and HDD or flash
memory for large program storage.
Effective, (but complex) hardware and software have
been developed to manage this memory hierarchy and
maximize its effectiveness.
47
N. B. Dodge 9/15

Computer Science
Exercise 2
1. Each small area of cache (say, 1K byte) represents a
much larger area (say, 1 Mbyte) in DRAM. If an
instruction, for example, is supposed to reside in a
given Mbyte of DRAM, the corresponding cache
extent is searched. Assume that, according to the
validity indicator, the correct instruction is NOT in
cache. What now?
2. Give simple definitions of the principles of temporal
and spatial locality.
48
N. B. Dodge 9/15

Computer Science
Computing in the Future

We discussed the evolution of computing
up to the present in Lecture #1.
Now lets talk about whats happening
today and in the near future.
Information shown here is from Intel,
The University of Texas, and various
other sources.
50
N. B. Dodge 9/15

Computer Science
Memory
In 2008, a completely new kind of circuit element (that
was predicted in 1971), the memristor, was developed.
Memristor circuit elements can retain a state (i.e.,
memory) even when power is off. They could eventually
replace flash EPROM, and perhaps DRAM. Thousand
Gbyte main memories are possible.
Memristors remember multiple states (not just ones and
zeros). Thus a memristor memory might eventually
remember like a human neuron, leading to neuraltype processors in the long term.
51
N. B. Dodge 9/15

Computer Science
Memristors: When?
HP says that it will have 100
TByte memristor drives by
2018.
The claim is that such drives
could be packaged in one
blade box for a total of 24
Petabytes!
There is no specific product
roadmap but HPs CTO has
talked confidently of HP
popping 100TB Memristor
drives into StoreServ arrays in
five years.
52
A memristor production slice. Expect

these production dates to be late, but
large memristor memories are possible
by 2020 or a bit later.
N. B. Dodge 9/15

Computer Science
Memory (3)
Another new memory type is Phase
Change Memory (PCM).
Heated to a high temperature, the device
produces an amorphous crystal
(disorganized structure), with high
conductivity. This is a 1.
To write a 0, the material is heated to a
lower temperature, creating an organized
crystal structure with low conductivity.
Still experimental. Can be read/written
about 100 million timesfar too low.
DRAM and flash memory can do 1-10
quadrillion cycles!).
53
N. B. Dodge 9/15

Computer Science
Memory (4)
Other new memory types:
Magnetic RAM (MRAM) Uses
tunneling resistance that
depends on the relative
magnetization of ferromagnetic
electrodes. Early work.
Resistive RAM (ReRAM)
Varies resistance according to
applied voltage. Nonvolatile,
low power, high density.
Production cost and reliability
are problems.
54
Only time will tell if these will challenge

DRAM.
N. B. Dodge 9/15

Computer Science
The CPU
Intel and AMD abandoned the GHz race in CPUs
years ago. In the early 2000s, Intel stated a goal of a
10 GHz CPU by 2010which clearly didnt happen.
Multiple CPUs became the performance enhancer.
The standard is now 4- and 6- core CPUs. Intel Xeon
server CPUs are 8-core.
With the new Broadwell architecture at the 14 nm
node, Intel is now said to be planning an 18-core CPU,
with 8- and 10-core performance CPUs as well. Speed
is inching upmaybe to 5 GHz in the late teens.
55
N. B. Dodge 9/15

Computer Science
Minimum Feature Size

Minimum feature size: the smallest dimension that
can be laid out on a chip (typically the gate).
Currently the minimum feature size for DRAM
memory is 28-23 nanometers. (one nanometer is one
billionth of a meter in length [109 meters]). Memory
size is forecast to go further down:
DRAM: 2015 2017 2019 2021 2023 2024 2025 2026
23 17.9 14.2 11.3 8.9 8.0 7.1 6.3
Currently the minimum feature size for DRAM

memory is 28-23 nanometers.
Note how far behind CPU chipmakers the memory
technology lags!
56
N. B. Dodge 9/15

Computer Science
CPU Technology Bounds Ahead

Intel and AMD currently
manufacture CPUs at the 22
nanometer node, and Intel is
sampling 14 nm products this
year. Intel will also start to
produce Skylake chips, the
successor to Broadwell.
Clearly, CPU chipmakers lead
the way in minimum feature
size.
AMD Laptop CPU Chip

57
N. B. Dodge 9/15

Computer Science
Three-D Technology
As feature size gets
smaller (22 nm14 nm),
3-D manufacturing
processes continue to
improve.
Generically referred to as
FINFET due to the 3-d
fin. Intel calls it trigate.
At right, comparison of
transistor sizes in 22 and
14 nm processes.
58
N. B. Dodge 9/15

Computer Science
Multi-Core Advance
Intel has announced a 72-core
CPU, Knights Landing.
Available in 2015, it will have
up to 16GB DRAM, up to
500GB/sec of memory
bandwidth, plus up to 384GB
of DDR4-2400 mainboard
memory. KL will use the 14nm
process. With a promise of 3 Intel Knights
Landing,
teraflops (double precision)
with 72 Pentium
per socket it will almost
cores with 64-bit
certainly be used to build some support.
monster x86 supercomputers.
59
N. B. Dodge 9/15

Computer Science
Stampede
New supercomputer at the University of Texas.
Uses several thousand Dell Zeus servers, each with
dual 8-core Intel Xeon processors.
Each Zeus server uses several Knights Corner chips, a
precursor to Knights Landing. Has 522,080 cores.
Knights Corner also uses modified Pentium-era cores.
Has 270 Tbytes (yes, that TERA bytes) of DRAM.
Has 14 Petabytes of storage memory.
Peak performance = 9,600,000,000,000,000 floating
point operations per second (9.6 petaflops).
Developers claim that exaflops are on the way.
60
N. B. Dodge 9/15

Computer Science
Intel: Sandy Bridge Skylake

What about the bread and butter Intel PC CPUs?
In the last few years, the Intel CPU family has gone from
Sandy Bridge (32 nm) Ivy Bridge (22 nm,
FinFET) Haswell (Optimized FinFET, 22 nm)
Broadwell (14 nm). The next generation is termed
Skylake (14 nm, new architecture, optimized FinFET).
Both Broadwell and Skylake will use much less power,
especially in the mobile-computing variants.
Haswell, (current-generation Intel CPUs) are all 22 nm.

Broadwell 14-nm chips are just appearing. Broadwell/
Skylake will be faster, lower power, have better graphics.
61
* Intel has odd CPU Family names, like Apple OS X updates Leopard, Snow Leopard, Lion.
N. B. Dodge 9/15

Computer Science
Tick-Tock
Intel is said to follow a Tick-Tock design strategy.
The Tick phase is shrinking the minimum feature size.
Tock is introducing a new microarchitecture. Thus:
62
Tick
Westmere (January, 2010)
Tock
Sandy Bridge (January, 2011)
Tick
Ivy Bridge (April, 2012)
Tock
Haswell (June, 2013)
Tick
Broadwell (October, 2014)
Tock
Skylake (Mid-2015)
Tick
Cannonlake (2016? 11 nm?)
N. B. Dodge 9/15

Computer Science
The End of Graphics Processors?

For high-performance graphics (e.g., video games), PC
gamers add one or more high-performance video cards.
With many-core chips, that may not be necessary. Many
of the cores may be used for graphics generation.
Better still, put GPUs on the chip! Interestingly enough,
Intel and nVidia signed a cross-licensing agreement in
2012, giving Intel access to nVidias GPU designs.
With the new Skylake family, it appears that graphics
processors are included as CPU cores. Graphics cards
may become unnecessary!
63
N. B. Dodge 9/15

Computer Science
The End of Moores Law?

Smaller nodes (that is, reducing minimum feature size) are getting
harder and harder to reach.
Some experts speculate that current production methods will not be
able to sustain CPU manufacturing below about 7 nm.
This means that we are probably coming to the end of the Moores
Law era, where the number of circuits per chip doubled every 18
months, and costs reduced by about half per transistor.
Will circuits continue to shrink in size?
Undoubtedly. But the current photographic (or shadow-mask)
manufacturing techniques will have to give way to other methods.
The only theoretical limits to the size of a switching element are
those related to the size of atoms and electronic wells!
64
N. B. Dodge 9/15

Computer Science
Windows 10Anything New?

Microsoft continues to lose ground in the mobile sector.
W10 appears to be a combination of Windows 8 and Windows 7.
The Start screen is back. Everybody hated W8s

desktop.
Users still have the option of Metro-style live tiles.
Virtual desktops. Unlike Linux, Windows has always
provided only one desktop. Now you can have as many as
you want.
Live tiles can be run as a window (app-like) on the
desktop. Alternatively, you dont have to use them.
65
N. B. Dodge 9/15

Computer Science
Windows (Continued)
Windows 10 has been free to early adapters via
download.
This probably will not remain true for long. However,
W10 may be offered at a steep discount to Windows XP
users to get them to ditch their 13-year-old operating
system. Maybe even a similar carrot for Vista users.
Other rumors state that W10 is the last Windows.
An OS that can be more easily updated (a la Apple)
may be next. It will be interesting to see Microsoft OS,
TNG (the next generation) and its features.
66
N. B. Dodge 9/15

Computer Science
Other Computer-Based State-of-the-Art Electronics

New 4K TVs are already
outfour times the
resolution of HD.
The other advance is
OLED, which has a much
sharper display than even
LED-backlit LCD displays.
Sixty-five inch TVs came
out recently75 and up are
coming.
Samsung now offers a 105inch, 4K, 3D LED TV with
a fully curved screen!
67
Samsung 105 4K TV
N. B. Dodge 9/15

Computer Science
3-D Printing
The development of 3D printers over the last few

years has been astonishing.
68
N. B. Dodge 9/15

Computer Science
Just About ANYTHING Can Be Printed!
69
N. B. Dodge 9/15

Computer Science
3-D Printing of Human Organs (!)
3D printers are being used more and more in bioengineering.

Above: Two 3-D printers in the act of printing biomatrices on
which organ cells can be attached to grow an organ.
70
N. B. Dodge 9/15

Lec21 PDF

Uploaded by

Copyright:

Available Formats

Lec21 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lec21 PDF

Uploaded by

Copyright:

Available Formats

The University of Texas at Dallas

Erik Jonsson School of Engineering and

Memory Management in Modern Computers

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas

Erik Jonsson School of Engineering and

Relative Speeds of the CPU and DRAM

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas

Erik Jonsson School of Engineering and

The Memory Speed/Cost Dilemma

However, there is a conflict in these requirements:

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas

Erik Jonsson School of Engineering and

Stating the Memory Management Problem

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas

Erik Jonsson School of Engineering and

Solving the Problem of Memory Management

High-speed electronic memory (DRAM, up to 64 Gbytes, fast)

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas

Erik Jonsson School of Engineering and

Lecture # 21: Memory Management in Modern Computers

Erik Jonsson School of Engineering and

The University of Texas at Dallas

We already know that registers are simply collections of D FFs.

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas

Erik Jonsson School of Engineering and

Random-Access Electronic Memory

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas

Erik Jonsson School of Engineering and

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas

Erik Jonsson School of Engineering and

As you saw on the previous slide, in terms of memory structure,

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas

Erik Jonsson School of Engineering and

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas

Erik Jonsson School of Engineering and

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas

Erik Jonsson School of Engineering and

Intel Core I7 Cache Structure

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas

Erik Jonsson School of Engineering and

Processor Layout of Single Intel Core i7 CPU

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas

Erik Jonsson School of Engineering and

Cache Location on Intel CPU

Lecture # 21: Memory Management in Modern Computers

The University of Texas at Dallas