Issues of existing computer hardware

for AI and ML

1 J. Joshua Yang
 Challenges future computing: four walls

Memory wall Dennard scaling

Architecture:

von Neumann Bottleneck

Device:

Moore’s law

Nature Electronics 1, 22–29 (2018, review)

2 J. Joshua Yang
 Background: Issues with classical computer hardware in the ‘big-data’ era
Architecture:
Classical computer 1. Separate processor and storage (Von
Neumann bottleneck)
2. Sequential process
3. Analog/Digital conversion
Transistors (static dopants):
• Electron manipulation: fast but volatile
• Doesn’t scale anymore: Moore’s law

Brain networks
Architecture:
Neurons, Synapses and networks
1. In-memory computing
2. Parallel comping Computing
Natural
3. Analog computing Accelerator
Intelligence
4. More neuroscience principles

Memristive Devices (mobile dopants):

• Ion manipulation: slow but nonvolatile
• Nanoionics can further scale down

Q. Xia, J. Joshua Yang, "Memristive crossbar arrays for brain-inspired computing” Nature Materials 18, 309 (2019)

3 J. Joshua Yang
 Post-CMOS era
Heterogeneous integrations
(development)

Existing technologies

Extended CMOS
Completely new concepts
(research)

New technologies
Memristor

(1) Emerging memory

(2) Emerging logic devices
(3) New architectures: Neuromorphic computing
ITRS 2013 (The International Technology Roadmap for Semiconductors) year

4 J. Joshua Yang
Emerging materials and devices are needed!

CMOS is the only choice to provide working systems now, but

may not be the ultimate solution with the most efficiency!

CMOS

Science 345, 668 (2014) (IBM Group)

IEEE Micro 38, 82 (2018) (Davies Group, Intel)
Nature 572, 106 (2019) (Shi group, Tsinghua)

100 W 100 W
1011 neurons 109 transistors
1014-15 synapses 10 transistors / synapse
Diffusive memristor è neuron 108 synapses
Drift memristor è synapse 105 neurons
Need ~ 107 chips
More efficient with emerging devices
inefficient with CMOS chips

5 J. Joshua Yang
 Devices: Mobilizing dopants/ions is the way to go!

CMOS:
• Manipulating electrons (fast, but volatile)
• Immobile dopants
• Static profile control
• Failed if dopant moves
• Software learning (devices not changing)
Different from biological systems, Intelligence?

Intelligent Materials and devices (IMD):

• Manipulating ions (slow but non-volatile)
• Mobile dopants
• Dynamic profile control (Electrically reconfigurable)
• Hardware learning (devices are changing)
Similar to biological systems, Natural intelligence!

However, we do not know how to precisely reconfigure dopants electrically yet!

6 J. Joshua Yang
 Architecture/algorithm: learning from the brain?

How can brain be so good with such lousy

devices ?
(slow, big, stochastic, inaccurate)

Probably because of its

smart
Architecture/Algorithm!

However, we do not know much about how

the brain works yet!

"Bridging Biological and Artificial Neural Networks with Emerging Neuromorphic Devices: Fundamentals, Progress, and Challenges" , Advanced Materials, 1902761, (2019, review).

7 J. Joshua Yang
History of memory devices

Courtesy of Prof. Han Wang

8 J. Joshua Yang
History of memory devices

Courtesy of Prof. Han Wang

9 J. Joshua Yang
Nonvolatile
Destructive reading

Courtesy of Prof. Han Wang

10 J. Joshua Yang
Different types of memory devices

Courtesy of Prof. Han Wang

11 J. Joshua Yang
 Courtesy of Prof. Han Wang

12 J. Joshua Yang
13 J. Joshua Yang
Amount of memory is increasing

Courtesy of Prof. Han Wang

14 J. Joshua Yang
 Courtesy of Prof. Han Wang

15 J. Joshua Yang
What’s in your iphone?
Memory is a big chunk!

Courtesy of Prof. Han Wang

16 J. Joshua Yang
Trend for memory: Nonvolatile memory becomes increasingly useful!

Courtesy of Prof. Han Wang

17 J. Joshua Yang
 Courtesy of Prof. Han Wang

18 J. Joshua Yang
 An example of a Nonvolatile
emerging memory technology

19 J. Joshua Yang
 Textbooks have been rewritten

“Electronic Devices and Circuit Theory” (11th Edition) “Introductory Circuit Analysis” (12th Edition)
Apr 30, 2012 January 15, 2010
by Robert L. Boylestad and Louis Nashelsky by Robert L. Boylestad

20 J. Joshua Yang
The Fourth fundamental passive circuit element

“Electronic Devices and Circuit Theory” (12th Edition)

2015, In press
by Robert L. Boylestad and Louis Nashelsky

21 J. Joshua Yang
 History of Memristor: Theory
Resistor – 1827
Georg Ohm

RESISTOR CAPACITOR
v=Ri q=Cv

Capacitor - 1745
Volta / von Kleist & van Musschenbroek
INDUCTOR Benjamin Franklin
φ=Li

Inductor – 1831
Michael Faraday
Joseph Henry

22 J. Joshua Yang
 History of Memristor: Theory
Introduction to Nonlinear Network Theory, NY: McGraw-Hill, 1969
1960’s – Mathematical Foundation for circuit theory of nonlinear systems Developed

RESISTOR CAPACITOR
dv = R di dq = C dv

i dq /dt = i q
dφ/dt = v

INDUCTOR “The Father of

dφ = L di
Non-Linear Circuit Theory”
f
But he also saw that there was a hole where
an obvious relation seems to be missing

23 J. Joshua Yang
 History of Memristor: Theory

“Memristor - the missing circuit element,”

v IEEE Trans. Circuit Theory 18, 507 (1971).

M is a “memory resistor” –
it is a nonvolatile resistive memory!
RESISTOR CAPACITOR
dv = R di dq = C dv Proved that no equivalent circuit to M
could be made using any R, C and L –
is a new and unique ‘basis function’
i dq /dt = i q
dφ/dt = v

INDUCTOR MEMRISTOR
dφ = L di dφ = M dq

24 J. Joshua Yang
 History of Memristor: Experiments
Nb/Nb2O5125nm/Bi

L. O. Chua and S. M. Kang, Proc. IEEE 64 209 (1976). W. R. Hiatt and T. W. Hickmott, Appl. Phys. Lett. 6, 106 (1965)

Research effort in searching such devices faded due to the rise of Si technology

Until the end of last centaury when the physical limit of Si technology is close.

25 J. Joshua Yang
 Brief introduction of memristive devices

Simple structure but rich materials, physics, chemistry, electrical issues

Metal 1
Metal 1

oxide
oxide V
Metal 2

Metal 2

26 J. Joshua Yang
The first functioning oxide Nanodevice

Cross-section
TEM
Nano-devices 1x17
BE Top electrode
TE
Localized region
oxide
50
n m
oxide
Conduction channel/filament

Bottom electrode

V 5nm

J. Joshua Yang et al., Nature Nanotechnology 3, 429 (2008)

27 J. Joshua Yang
 Image and signature of memristors

Nano-devices 1x17
BE 200
TE

100

Current ( uA )
50
n m
oxide
0

-100

-200-2
-1 0 1 2
Voltage ( V )
V

J. Joshua Yang et al., Nature Nanotechnology 3, 429 (2008)

28 J. Joshua Yang
 Promises
Promises 85ps ON/OFF switching
Challenges
1. Speed (85ps)
Speed (85ps) • Mechanism
2. Scalability (2nm)
3. Multilevel (>64) 1. Variability

4. Stackability (>8 layers) Nonvolatile

2. Nonlinearity
5. CMOS compatibility memory
(selector)
6. Non-volatility (>10years)
7. Non-destructive reading 3. Synaptic dynamics
Neuromorph
8. Low cost
ic computing
9. …… Choi, et al, Adv. Funct. Mater. 26, 5290 (2016).

29 J. Joshua Yang
 Promises

Promises 2nm x 2nm memristor crossbar

Challenges
1. Speed (85ps) • Mechanism

2. Scalability
Scalability(2nm)
(2nm)
1. Variability
3. Multilevel (>64)
4. Stackability (>8 layers) Nonvolatile
2. Nonlinearity
5. CMOS compatibility memory
(selector)
6. Non-volatility (>10years)
7. Non-destructive reading 3. Synaptic dynamics
Neuromorph
8. Low cost ic 35
computing
Pi et al., Nature Nanotechnology 14, (2019).
9. …… (Q. Xia)

30 J. Joshua Yang
 Promises

Promises Challenges
>64 resistance levels
1. Speed (85ps) • Mechanism
2. Scalability (2nm)
3. Multilevel
Multilevel(>64)
(>64) 1. Variability

4. Stackability (>8 layers)

2. Nonlinearity
5. CMOS compatibility
(selector)
6. Non-volatility (>10years)
7. Non-destructive reading 3. Synaptic dynamics
Neuromorph
8. Low cost
ic computing
9. …… Li et al., Nature Electronics 1, 52 (2018).

31 J. Joshua Yang
 Promises

Promises Challenges
3D integration of memristor on CMOS
1. Speed (85ps) • 8 Layers
Mechanism

2. Scalability (2nm)
1. Variability
3. Multilevel (>64)
4. Stackability (>8(8layers)
Stackability layers) Nonvolatile
2. Nonlinearity
5. CMOS compatibility memory
(selector)
6. Non-volatility (>10years)
7. Non-destructive reading 3. Synaptic dynamics
Neuromorph
8. Low cost ic computing
9. …… Lin et al., Nature Electronics 3, 225 (2020)
(Q. Xia)

32 J. Joshua Yang
 (1) Emerging Memory:
exponential growth in demand

Non-volatile Memory Gigabyte Growth Trends

“Cloud” is really
memory and
Storage

https://www.grandviewresearch.com/industry-analysis/next-generation-non-volatile-memory-market

33 J. Joshua Yang
 (1) Emerging Memory:
RRAM – fast, non-volatile, low energy, Ultra-Dense

0 nm
3
m
2µ
Switching materials
(e.g. TiO2)

Smaller cell (scalability) and 3D (stackability) lead to ultra-high density!

34 J. Joshua Yang
 (1) Emerging Memory:
Memristor – fast, non-volatile, low energy

Flash
Memristor PCM STTRAM SRAM DRAM HDD
(NAND)

Reciprocal density (F2 ) <4 4-16 20 - 60 140 6-12 1-4* 2/3

Energy per bit (pJ) 0.1–3 2-25 0.1-2.5 0.0005 0.005 0.00002 1–10x10 9

Read time (ns) <10 10-50 10-35 0.1-0.3 10 100000 5–8x10 6

Write time (ns) ~10 50-500 10-90 0.1-0.3 10 100000 5–8x10 6

as long as
Retention years years years voltage << second years years
applied

Endurance (cycles) 10 12 10 9 10 15 >10 16 >10 16 10 4 10 4

prototype and … … commercialized technologies

J. Joshua Yang et al., Nature Nanotechnology 8, 13 (2013)

35 J. Joshua Yang
 (2) Emerging logic:
Type I: Memristor/CMOS hybrid logic – FPGA

Q. Xia et al., Nano Letters., 9, 3640 (2009)

36 J. Joshua Yang
 (2) Emerging Logic:
Type II: implication logic

Bertrand Russell

Material Implication + FALSE operation

= universal computing

OR, AND, NOT

“Material implication”
The forth fundamental logic operations

37 J. Joshua Yang
 (2) Emerging Logic:
Type II: implication logic based NAND

Computation Naturally realized with nano-

crossbars:
Two Memristors à implication logic
Three Memristors à NAND logic

Advantages:
• Denser
• Non-volatile logic

P Q S

RG

s ß q IMP (p IMP 0) = p NAND q

J. Borghetti et al., Nature 464, 873 ( 2010)

38 J. Joshua Yang
(3) Computing with Memristive devices
Natural Intelligence Original
Improved speed-energy
Gaussian Disk
efficiency
Average

Artificial
Synapses
Nature Comm. 8, 752 (2017)
Nature Materials 16, 101 (2017)
Science 364, 570 (2019)
Laplacia(2020)
Sci. Adv. 6, eaba6173 LoG Prewitt (Horizontal) Sobel (Horizontal)
Original Gaussian Disk
Nature Comm. 11,n 1861 (2020)
64 x 64 1T1R array

Artificial ML
Memristor Accelerators
Neurons
Nature Electronics 1, 137 (2018) array Nature Electronics 1, 52 (2018)
Nature Comm. 9, 3208 (2018) Nature Comm. 9,Laplacia
2385 (2018)
Nature Comm. 11, 51 (2020) LoG Prewitt (Horizontal)
Adv. Mater. 29, 1705914
n (2018)
Nature Nanotechnology 15, 776 (2020) Nature Machine Intelligence 1, 49 (2019)
Nature Electronics 2, 115 (2019). 64 x 64
Nature Machine Intelligence 1, 434 (2019)
Example: Sensing System For Robotics Nature Electronics 3, 409 (2020)
Nature 577, 641 (2020)
Intelligent
Systems
Nature Comm. 8, 882 (2017). Nature Materials 16, 396 (2017).
Nature Comm. 8, 15666 (2017). Nature Comm. 9, 417 (2018)
Adv. Mater. 29, 1604457 (2017). Nature Electronics 1, 130 (2018).
Adv. Mater. 31, 1902761 (2019) Nature Electronics 1, 548 (2018).
Nature Materials 18, 309 (2019). Nature Nanotechnology 14, 35 (2019).
Nature Electronics 3, 225 (2020) Nature Review Materials 5, 173 (2020)

39 J. Joshua Yang
Computing with memristive networks

With little Bio-inspiration:

Just a resistive network

for computing acceleration (parallel, analog, in-memory)
With some Bio-inspiration:
a reprogrammable network
for pattern recognition (after supervised learning)

With more Bio-inspiration:

a bio-realistic neural network

for pattern classification (after unsupervised learning)

40 J. Joshua Yang
Vector x Matrix multiplication accelerator
•Computation intensive tasks (many multiplication and addition steps)
•Only one step in dot-product engine

1. In-memory computing
2. Parallel comping
3. Analog computing

Computing by Ohm’s law

and Kirchhoff's current law
>100x improvement in speed & energy efficiency!
Hu et al., Adv. Mater. 29, 1705914 (2018).

41 J. Joshua Yang
Hardware demonstration of accelerator

Li et al., Nature Electronics 1, 52 (2018).

(Cover image of inaugural issue)

42 J. Joshua Yang
Image Processing (experimental demo)

Image Compression Using Analog Memristor Array (DCT) (20:3 compression ratio)
Original image Software processed Memristor array processed (Experimental)

Li et al., Nature Electronics 1, 52 (2018).

43 J. Joshua Yang
Computing with memristive networks

With little Bio-inspiration:

Just a resistive network

for computing acceleration (parallel, analog, in-memory)
With some Bio-inspiration:
a reprogrammable network
for pattern recognition (after supervised learning)

With more Bio-inspiration:

a bio-realistic neural network

for pattern classification (after unsupervised learning)

44 J. Joshua Yang
The key idea in brain-inspired computing

The brain is a complex

network of neurons, connected
by synapses

http://www.macs.hw.ac.uk/~dwcorne/

45 J. Joshua Yang
The key idea in brain-inspired computing

When neurons are active, they

send signals to others.

http://www.macs.hw.ac.uk/~dwcorne/

46 J. Joshua Yang
The key idea in brain-inspired computing

A neuron with lots of ‘strong’

active inputs will become active.

http://www.macs.hw.ac.uk/~dwcorne/

47 J. Joshua Yang
The key idea in brain-inspired computing

And, when connected neurons

are active at the same time, the
link between them gets stronger

Learning/training:

Fire together wire together

(Hebbian learning rule)
Spike-timing-dependent plasticity
(STDP)
http://www.macs.hw.ac.uk/~dwcorne/

48 J. Joshua Yang
The key idea in brain-inspired computing

So, suppose these neurons happen

to be active when you see a fluffy
animal with big eyes, small ears and a
pointed face

http://www.macs.hw.ac.uk/~dwcorne/

49 J. Joshua Yang
The key idea in brain-inspired computing

So, suppose these neurons happen

to be active when you see a fluffy
animal with big eyes, small ears and a
pointed face

… and suppose your mom then

says “Cat”, which excites this
additional neuron.

http://www.macs.hw.ac.uk/~dwcorne/

50 J. Joshua Yang
The key idea in brain-inspired computing

Links will then strengthen between

the active neurons

Inferencing with trained network

So when you see a similar animal

again, this neuron will probably
Automatically be activated, helping
you classify it.

A slightly different group of neurons

will respond to dogs …

In this way, the neural networks learn

gradually from examples.
http://www.macs.hw.ac.uk/~dwcorne/

51 J. Joshua Yang
Beyond accelerator: Neural Networks capable of learning

Li et al., Nature Communications 9, 2385 (2018).

52 J. Joshua Yang
2-layer Neural Network for in-situ Online Training

Li et al., Nature Communications 9, 2385 (2018).

53 J. Joshua Yang
Powerful applications, but the training is demanding…

Nature Electronics 1, 52 (2018) Nature Comm. 9, 2385 (2018) Nature Electronics 3, 225 (2020)

Nature Machine Intelligence 1, 49 (2019) Nature Electronics 2, 115 (2019).

Nature Machine Intelligence 1, 434 (2019)

54 J. Joshua Yang
Computing with memristive networks

With little Bio-inspiration:

Just a resistive network

for computing acceleration (parallel, analog, in-memory)
With some Bio-inspiration:
a reprogrammable network
for pattern recognition (after supervised learning)

With more Bio-inspiration:

a bio-realistic neural network

for pattern classification (after unsupervised learning)

55 J. Joshua Yang
 The origin of the timing in bio-synapse: ‘programming’
Very Large Impulse from upper axon terminal à open NMDA receptor, Ca2+ diffuses
through à activate some bio-activities à increasing both the number and conductance of
AMPA receptorsà synapse enhanced – long term plasticity.

Again, the Ca2+ dynamics controls the synaptic weight changes!

http://sites.sinauer.com/neuroscience5e/animations08.02.html

56 J. Joshua Yang
Similar Diffusion dynamics in artificial devices

Ag: SiNxOy

Z. Wang et al., Nature Materials 16, 101 (2017)

57 J. Joshua Yang
Snapshots

58 J. Joshua Yang
Experimental results: Short-term and long-term synaptic plasticity

Ag memristor
Paired pulse facilitation/depression

TaOx memristor
Ag memristor

Conductance

Spiking-timing dependent plasticity

Presynaptic

TaOx memristor

Ag memristor
Postsynaptic

Z. Wang et al., Nat. Electron. 1, 137 (2018).

59 J. Joshua Yang
Artificial Neurons: role of ion dynamics in bio-neurons

Action
Potential

Repolarization

Depolarization

ATP pump keeps

neuron in equilibrium

© 2012 Pearson Education, Inc.

60 J. Joshua Yang
Emulating an Ion channel in a Neuron

61 J. Joshua Yang
Experimental results: leaky integrate-and-fire neuron

Ion Channel

Conductance
Ag memristor

Plasma Membrane

Z. Wang et al., Nat. Electron. 1, 137 (2018).

62 J. Joshua Yang
The 1st fully memristive spiking neural network
8×8 Ta/HfOx
memristor synapses

8 Ag memristor neurons
(with external capacitors)

Z. Wang et al., Nat. Electron. 1, 137 (2018).

63 J. Joshua Yang
Memristor spiking neural network: a self-organizing map

4×4 Input Convolution layer

(e.g. “U, M, A, S”) (Fixed)

Z. Wang et al., Nat. Electron. 1, 137 (2018).

64 J. Joshua Yang
Experimental results: Unsupervised learning with the 1st fully memristive NN
Convolution layer Dense layer
(Fixed) (physics driven automatic evolution)

Z. Wang et al., Nat. Electron. 1, 137 (2018).

65 J. Joshua Yang
Summary
Materials and Devices: beyond CMOS

CMOS

Science 345, 668 (2014) (IBM Group)

IEEE Micro 38, 82 (2018) (Davies Group, Intel)
Nature 572, 106 (2019) (Shi group, Tsinghua)

100 W 100 W
1011 neurons 109 transistors
1014-15 synapses 10 transistors / synapse
108 synapses
Diffusive memristor è neuron
Drift memristor è synapse
105 neurons
Need ~ 107 chips

More efficient with emerging devices inefficient with CMOS chips

66 J. Joshua Yang
 Devices: Mobilizing dopants/ions is the way to go!
CMOS:
• Manipulating electrons (fast, but volatile, good for logic)
• Immobile dopants
• Static profile control in the fabrication stage
• Failed if dopant moves
• Software learning (devices not changing)
Different from biological systems, Intelligence?

Intelligent Materials and devices (IMD):

• Manipulating ions (slow but non-volatile, good for memory)
• Mobile dopants
• Dynamic profile control (Electrically reconfigurable)
• Hardware learning (devices are changing)
Similar to biological systems, Natural intelligence!

However, we do not know how to precisely reconfigure dopants electrically yet!

67 J. Joshua Yang
 Architecture/algorithm: learning from the brain?
How can brain be so good with such lousy
devices ?
(slow, big, stochastic, inaccurate)

Probably because of its

corresponding
Architecture/Algorithm!

However, we do NOT know much

about how the brain works yet!

"Bridging Biological and Artificial Neural Networks with Emerging Neuromorphic Devices: Fundamentals, Progress, and Challenges" , Advanced Materials, 1902761, (2019, invited review).

68 J. Joshua Yang
 Outlook: Roadmap
W.G. Kim et al., SK-hynix
J. Joshua Yang et al., HP Labs
VLSI Technology Symposium
(2014) 138-139

DRAM

Flash
RAM
Flash
Hard Disk
Solid State Disk Optical disk
Floppy Disk

Chip Storage Class Universal Analog and Neural

development Memory Memory Computing

Time

69 J. Joshua Yang
How materials principles used in emerging
devices?

Examples

70 J. Joshua Yang
 Simplified picture of switching mechanism: TiOx

Where: interface;
What: oxygen vacancies (2+, n-dopant); Schottky-like: OFF

How: drift.
w

metallic Φb

OFF
ON
Pt TiO2

V
Ohmic-like: ON

w
TiO2-X
Φb

Ti TiO2
J. Joshua Yang et al., Nature Nanotechnology 3, 429 (2008)
J. P Strachan et al., Advanced Materials 22, 3573 (2010)
F. Miao, J.P. Stranchan, J. JoshuaYang* et al., Adv. Mater. 23 5633 (2011) Vo2+ n-type dopant

71 J. Joshua Yang
 Switching is a result of ionic motion under the
combined effect of electric field and Joule heating

Field dominating thermal dominating

Bipolar Bipolar Nonpolar Nonpolar

TiOx TaOx NiOx NbO2

Nonlinear Linear Bi-stable Threshold
IV IV

Current
Current
Current

Current
Voltage
Voltage Voltage Voltage

Electric e- Electric Ec
Concentration Temperature
potential potential gradient
gradient Eb
gradient gradient
Et
e- e- - Ef
e- e
Cathode
J. Joshua Yang et al., Nature Nanotechnology 8, 13 (2013) Ef
e- 72 J. Joshua Yang
e-
 Nonlinear
IV
thermal dominating
IV

Current
Current
Mechanisms: four Ionic transport mechanisms
Bipolar Nonpolar Nonpolar
Voltage Voltage
Drift Electromigration
Electric e- Electric
Concentrati
potential potential
gradient
Linear gradient Bi-stable
gradient
IV e- -
e-
Current

Current
e- e
e-
e-

Voltage Voltage
Fick Diffusion Thermophoresis
-
Electric Temperature Ec
Concentration
potential gradient
gradient Eb
gradient

Ef
Cathode

Ev Oxid

J. Joshua Yang et al., Nature Nanotechnology 8, 13 (2013)

73 J. Joshua Yang
 Possible Electron transport mechanisms

(1) Schottky emission

(2) Fowler-Nordheim (F-N) tunneling
(3) Direct tunneling
(4) tunneling from cathode to traps
(5) emission from trap to conduction band
(6) F-N - like tunneling from trap to conduction band
(7) trap to trap hopping or tunneling
(8) tunneling from traps to anode.
S.M. Yu et al., APL 99, 063507 (2011)

74 J. Joshua Yang
Nanodevice family: M/I/M with two interfaces
Pt 10
TiO2 I 0 I
TiO2-X Measured
V
V -10
-1.5 1.5 V
-V
TiO2-X
Complementary

+V

Intrinsic diode
Pt

5 1.5
switching

1.0 Measured
0

+V
-V
+V
0.5

-V
-5 Measured
0.0
-4 -2 0 2 -2 0 2

I I

V
V

75 J. Joshua Yang