computing with unreliable nanodevices - gdr...
TRANSCRIPT
![Page 1: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/1.jpg)
Computing with
Unreliable Nanodevices
Damien Querlioz
Chargé de recherche CNRS, https://sites.google.com/site/damienquerlioz/
Centre de Nanosciences et de Nanotechnologies
(ex-Institut d’Electronique Fondamentale)
Univ. Paris-Sud, Université Paris-Saclay, CNRS, Orsay
1
GDR BIOCOMP Summer School
Roscoff, June 2017
![Page 2: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/2.jpg)
Microelectronics has achieved a lot
• Right now we can now incredible things…
2
High Perf graphic
processor (GPU)
Nvidia TESLA K10
>7 billions transistors
>Tflops on a single
chip !
High Perf
microprocessor (CPU)
Intel Xeon E7
>2 billions transistors
100W
Mobile CPU and GPU
Based on ARM designs
3D gaming with <1W !
![Page 3: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/3.jpg)
How this happened
3
1
10
100
1000
104
105
106
107
1980 1990 2000 2010
Nb transistors
(/1000)
Horloge (MHz)
Puissance (W)
CPU Intel (tendances)
Source : Intel, Wikipedia, K. Olukotun
But things are changing
On a single chip, all transistors work
![Page 4: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/4.jpg)
Microelectronics « crisis »
• Since transistors have entered the nanometer regime, transistor scaling brings little benefit
• Very difficult to make nanotransistors reliable
4
The end of the traditional model of
microelectronics
April 2015
![Page 5: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/5.jpg)
« More than Moore » research
• Nanotechnology brings new devices to CMOS that can give it new features
Novel functions: memory effects, sensor, energy harvesting…
5
More than Moore comes with challenges
Novel memories
CEA LETINEMS sensors
![Page 6: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/6.jpg)
Electronic nanodevices : the good and the challenge
• Nanoelectronics’ devices are amazing
– Compact, low power
– Novel functions: memory effects, sensor, energy harvesting…
• But…
– Variability
– Noise, Faults
• Few products with nanoelectronics!
6
Can we design for technology's imperfections?
CEA LETI
Borghetti et al, Nature 464 (2010)
![Page 7: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/7.jpg)
Biology: the nanoelectronics champion?
Controlled current
source
Ion channel
~5nm
~500nm
Brains are extremely energy efficient based on
nanodevices. Can it be an inspiration?
Connection, nonlinear
dynamics, memory
Synapses
Neuron compute using nanoelectronics/nanoionics devices
7
![Page 8: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/8.jpg)
Biology’s nanoelectronics
• Biological « nanodevices » are not ideal electron devices
Example of ion channel (controlled current source)
Current is pure telegraph noise!
And in a neuron, only ~100 ion channels open at the same time (Schneidman, 1998)
Total ion channel current fluctuates by 100 100 = 10% !
8
Hille et al, 2001
Can we take some inspiration from the way biology
manages to compute with “imperfect” devices?
Curr
ent
5 pA
0 pA
![Page 9: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/9.jpg)
Question of the class
• We struggle with unreliable nanodevices, whilebiology shows us it is possible to strive with them
9
This class:
Can we also compute reliably with awesome but imperfect
nanodevices?
![Page 10: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/10.jpg)
Outline of the class
• The reliability issue of nanodevices
• Computing with errors: “Detect and correct”
• Approximate computation
• Computing with noise
• Toward bioinspired computing
10
![Page 11: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/11.jpg)
Outline of the class
• The reliability issue of nanodevices
• Computing with errors: “Detect and correct”
• Approximate computation
• Computing with noise
• Toward bioinspired computing
11
![Page 12: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/12.jpg)
Nanodevices are imperfect
• Variability
• Noise / unpredictability
12
First, let’s look at the case of CMOS
![Page 13: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/13.jpg)
NanoCMOS: high variability becomesintrinsic
• Only a few dozens of dopants in the channel
13
LG
Gate
DrainSource
W
Variation in number of dopants causes unavoidable
mismatch between transistors
![Page 14: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/14.jpg)
Fully Depleted SOI (FDSOI)
• ST Microelectronics since 28nm
• Now Samsung and Globalfoundries
14Weber et al, IEDM 2008
Source Drain
Gate
OXIDE
Channel
Silicon on Insulator wafer
![Page 15: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/15.jpg)
Fully Depleted SOI (FDSOI)
15Weber et al, IEDM 2008
Source Drain
Gate
OXIDE
Channel
Channel is
created physically
- Low leakage
- No need for
channel doping!
![Page 16: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/16.jpg)
The other option : FinFET
• Intel: since 22nm
• Also TSMC, Samsung, Gloablfoundries…
16
Burried channel surrounded
by gate!
Undoped channel
![Page 17: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/17.jpg)
These solutions still have limitations
• New variability issues when ultrascaled
• (gate material…)
17
How about true nanoelectronics ?
![Page 18: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/18.jpg)
Carbon nanotubes FETs 1/2
• Variability issue of carbon nanotubes
18
Shulaker et al ,
Nature 2013
Huge intrinsic variety of carbon nanotubes
Sorting techniques inaccurate!
![Page 19: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/19.jpg)
Carbon nanotubes FETs 2/2
• Low frequency noise order of magnitudes higher than bulk materials!
• Noise issue:
– Surface vs. volume
19
Bulk case Nano case
![Page 20: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/20.jpg)
Stochastic switching in Spin Torque Magnetic Memory
• Switching fundamentally stochastic due to basic magnetism physics (no switching without thermal noise)
20
Measurements on an in-
plane magnetization
structure (T. Devolder,
IEF/Univ. Paris-Sud)
V0 : arbitrary unit (<1V)
Vincent et al, IEEE Trans Biomed Circuits and Systems, 9, p. 166, 2015
![Page 21: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/21.jpg)
Resistive memory (memristors) 1/2
• Memristors: memories that switch due to atomic effects
21
![Page 22: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/22.jpg)
Resistive memory (memristors) 2/2
• Considerable variability and impredicatbility!
22
MemristorsHP Labs, Borghetti et al, Nature 464 (2010)
![Page 23: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/23.jpg)
How do we normally deal with imperfections?
• Conventional microelectronics:
–Design for Worst Case
23
![Page 24: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/24.jpg)
Example: setting circuit’s clock frequency
• 𝑓 <1
𝑡𝑚𝑎𝑥
• 𝑡𝑚𝑎𝑥 : time of the longest delay that can happen within a clock cycle, assuming all the transistors are the worst possible ones!
24
![Page 25: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/25.jpg)
Problem with worst case design
• It largely wastes the potential of a technology
• Significant issue with modern CMOS…
25
… and maybe not sustainable at all
with nanodevices!!!
![Page 26: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/26.jpg)
Two big strategies for « Better-than-worst-case » design
• Detect and correct errors
– Example here: RAZOR
• Accept an approximate result
– The approach of Biology?
26
![Page 27: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/27.jpg)
Outline of the class
• The reliability issue of nanodevices
• Computing with errors: “Detect and correct”
• Approximate computation
• Computing with noise
• Toward bioinspired computing
27
![Page 28: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/28.jpg)
The RAZOR system: idea
• I want to design a low power system at e.g. 100MHz
• Usually: I choose lowest supply voltage so that circuit alwaysworks at 100MHz in the worst case situation
• Here: I choose supply voltage so that circuit typically works at100MHz
And I find a way to detect if an operation did not have time to finish so that it can be flushed
28
![Page 29: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/29.jpg)
The RAZOR system
29Ernst et al ASPDAC 2005
![Page 30: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/30.jpg)
Results
30Ernst et al ASPDAC 2005
![Page 31: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/31.jpg)
Outline of the class
• The reliability issue of nanodevices
• Computing with errors: “Detect and correct”
• Approximate computation
• Computing with noise
• Toward bioinspired computing
31
![Page 32: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/32.jpg)
Approximate computing
• Let’s think about it…
• Do we really need absolutely exact result for everything?
• The cost of exactness
32
![Page 33: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/33.jpg)
A current trend toward approximate
33
IEEE Spectrum 2015
Computer 2014
![Page 34: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/34.jpg)
How are real numbers coded in electronics?
• Fixed Point representation
• Floating Point representation
34
Biology certainly does not use
Floating Point
![Page 35: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/35.jpg)
Modern application: Neural networks with Fixed Point
• Neural networks usually simulated with 32 bits or 64 bits floating point (GPU or CPU)
• Inference also works in low precision Fixed Point!
35
![Page 36: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/36.jpg)
Approximate neural network
36
Stromatias et al, Front Neurosci 2015
Machine learning applications are especially adapted
to approximate computing
Lin et al, ICML 2016
![Page 37: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/37.jpg)
Going further: Accepting incorrect least significant bits
37
Example with CMOS
Ernst et al ASPDAC 2005
Very significant energy saving (V²) if you accept errors
![Page 38: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/38.jpg)
Need for associated ecosystem
38
![Page 39: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/39.jpg)
Outline of the class
• The reliability issue of nanodevices
• Computing with errors: “Detect and correct”
• Approximate computation
• Computing with noise
• Toward bioinspired computing
39
![Page 40: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/40.jpg)
Can we give up determinism?
• Another idea: stochastic computing
• How to compute product with stochastic bit streams? (like Gaines)
• Precision as a function of observation time
40
![Page 41: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/41.jpg)
Other examples of stochastic computing
• Sum
• Square
41
The big challenge of correlations
AND AND
Delay
P(X)P(Z)=P(X) !!!
P(X)P(Z)=P(X)²
P(X)
P(Y)
P(C)
P(Z)=P(C)P(X)+(1-P(C))P(Y)MUX
011011011011
011011
101101
001001
![Page 42: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/42.jpg)
Stochastic computing can be very adapted to inference tasks
• Idea: encode probabilities by probabilities• A single C element (8 transistors) implements Bayes rule!
• They can be cascaded:
• But brings challenges too! (temporal correlations)
42Friedman et al, IEEE TCAS I, 2016; IJAR 2017
Spam detector
𝑃(𝑍) =)𝑃(𝑋)𝑃(𝑌
𝑃(𝑋)𝑃(𝑌) + (1 − 𝑃 𝑋 )(1 − 𝑃 𝑌 )
![Page 43: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/43.jpg)
Example of spam classification
43
A Do you want to get pizza for lunch?
B Can you please check my stochasticsimulations?
C If you want to earn a fortune, send a $100 check to Nigeria and we will transfer $10,000 to your account.
D My weekly commute to Nigeriaincludes a transfer in Morocco. I will check if my flight is on schedule- if so, do you want to get pizza when I arrive?
E There is a $10 fee for all checktransfers.
A promising lead, lots of work necessary
Friedman et al, IEEE TCAS I, 2016; IJAR 2017
![Page 44: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/44.jpg)
Crazier ideas: stochastic resonance
• Seen in neurons in vitro
• Some evidence suggests it may be used by the brain
44
Similarly, some circuits can have behavior improved by noise!
First understood in the 80’s for quaternaryglaciations [C.Nicholis]Now also observed in physics and biology
Idea: increasing a system’s sensitivity to small inputs by adding noise
![Page 45: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/45.jpg)
The canonical theory of stochatic resonance
45
A hypothetical particle that evolves along
Bistability Subthreshold
driveNoise (D)
Periodical part of the response
Periodical response has a maximum as
a function of noise
![Page 46: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/46.jpg)
46
thermal noise
DE P AP
Small size → low barrier → easy switching
→ high frequency oscillations
Same basic technology as Spin Torque MRAMEnergy barrier ΔE depends on material and volume of the free layer
CoFe 2.5 nm
Ru 0.85 nm
CoFeB 2.4 nm
MgO 1.0 nm
CoFeB 1.7 nm
Superparamagnetic Tunnel Junction
![Page 47: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/47.jpg)
47
0 100 200 300 400 500
150
175
200
Time (ms)
R (
Ohm
s)
0
50
100
150
Co
un
t
Anti-Parallel
State
0 10 20 30 40 50 60 700
50
100
150
Parallel
State
Dwell-times (ms)
<AP> = 8.5 ms
<P> = 10.1 ms
AP
P
𝐹0 =1
<𝜏𝑃>+<𝜏𝐴𝑃>= 53.8 𝐻𝑧 F0 : natural frequency with no source
Poisson process
Measured at low DC current:IDC
Poissonian Statistics of Switching
Locatelli et al, Phys Rev Appl 2014; Mizrahi et al, IEEE Trans Mag 2015
![Page 48: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/48.jpg)
Experiment: noise controls frequency and phase locking Experiment: Noise Controls Frequency and Phase Locking
48
𝐹 =1
< 𝜏𝐴𝑃 > +< 𝜏𝑃>
Number of oscillations per
second
Vac = 63 mV while
Vc = 235 mV @0K
ΔE = 22.5 kBT
Natural frequency ≈ 0.1 Hz
Thermal noise (room
temperature)
+
White electrical noise
0 5 10 15 20 25 30 35 400
10
20
30
40
50
60
70
0 50 100 1500 50 100 150 0 50 100 150 200
1
-Vac
RP
Fre
quen
cy (
Hz)
Electrical noise (mV)
1
Fac
2
3
2 3RAP
+Vac
Re
sis
tance
sta
te
Time (ms)Time (ms)
Time (ms)
D
rive
vo
ltage
Mizrahi et al., Scientific Reports 2016
![Page 49: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/49.jpg)
Effect of Drive Amplitude and Frequency
49
Fac = 50 Hz Vac = 82 mV
Natural frequency ≈ 0.1Hz
0 20 40 60 80 100
0
10
20
30
40
50
101
102
103
104
0
10
20
30
40
50
60
E
lectr
ical nois
e (
mV
)
Drive voltage (mV)
Ele
ctr
ical nois
e (
mV
)
Drive frequency (Hz)
Model Experiments
→ Synchronization possible at broad ranges of amplitudes
and frequencies
SyncSync
Boundaries synchronization
Mizrahi et al., Scientific Reports 2016
![Page 50: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/50.jpg)
Minimum Energy Required for Synchronization
50
∆𝐸 = ∆𝐸0
𝐷2
𝐷02
𝑅 =𝑅𝐴
𝜋𝐷2
4𝑉𝑐 = 𝑐𝑡𝑒
Out of
plane
magnetic
tunnel
junction
• Optimal junction diameter (natural frequency ≈ drive frequency)
• Adding electrical noise is energy efficient
• Emin < 10-13 J
Sato et al., APL, 20140 5 10 15 20 25 30
10-14
10-13
10-12
10-11
10-10
10-9
10-8
10-7
10-6
10-5
10-4
Fac
= 10MHz
Fac
= 100kHz
En
erg
y (
J)
Diameter (nm)
Thermal + electrical noise
Thermal noise only
Fac
= 100Hz
Mizrahi et al., Scientific Reports 2016
![Page 51: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/51.jpg)
Industrial challenge of Better than Worst Case
• How does industry feel about better than worst case design?
51
![Page 52: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/52.jpg)
Outline of the class
• The reliability issue of nanodevices
• Computing with errors: “Detect and correct”
• Approximate computation
• Computing with noise
• Toward bioinspired computing
52
![Page 53: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/53.jpg)
Neuroinspiration
• Biology does not work with digital logic, uses approximate and redundant coding, can perform advanced computation, at low power
• A supercomputer (MW) cannot do what a brain does (20W)
53
vs.
![Page 54: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/54.jpg)
Main ideas
• Nanodevices with a lot of functionality
• Massive parallelism
• Slow, low power operations
54
![Page 55: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/55.jpg)
An example:Synaptic computation
• Synapse = self-adapting connection between two basic computing units (can change electrical resistance)
• Synapse = transmission + learning
With CMOS, difficult to cointegrate
logic/memory
But natural with nanoelectronics
Arthur,
NIPS 2006
55
![Page 56: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/56.jpg)
Bioinspired nanoelectronics
• Some nanodevices encompass memory (e.g. spin torque magnetic memory)
• Perfect element for implementing synapses!
• Neural networks have excellent resilience to their elements’ imperfections
56
BIOCOMP: G. W. Burr’s lecture
![Page 57: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/57.jpg)
Conclusions
• Nanodevices tend to be less reliable than traditional electron devices
• Worst case design may not be sustainable with nanodevices
• Alternate computing approaches that detect errors or accept approximate results may better benefit from nanodevices qualities
• Groundbreaking ideas (computing with noise or like biology) may allow true reinvention of computing with nanodevices
57
![Page 58: Computing with Unreliable Nanodevices - GDR BioCompgdr-biocomp.fr/wp-content/uploads/2017/07/Damien-Querlioz.pdf · Computing with Unreliable Nanodevices ... Intel Xeon E7 >2 billions](https://reader033.vdocuments.us/reader033/viewer/2022051601/5ab6e8f97f8b9a1a048e569a/html5/thumbnails/58.jpg)
58
Thank you for your attention!