dac skytalk and - armkeil.blob.core.windows.net€¦ · $10,000 $100,000 $1,000,000 $10,000,000...
TRANSCRIPT
DAC SKYtalk
Design and
Manufacturingin 2030
Greg YericFellow, Arm Research
Guidelines (1)
• Please keep the default font size for main lines at 28pt (or 26pt)• And use 24pt (or 22pt) font size for the sub bullets
• Use the default bullet style and color scheme supplied by this template
• Limited the number of bullets per page
• Use keywords, not full sentences
• Please do not overlay DAC logo
• Check the page numbering
© Arm 20202
$10,000
$100,000
$1,000,000
$10,000,000
$100,000,000
$1,000,000,000
1960 1970 1980 1990 2000 2010 2020 2030 2040
Moore’spaper
Sco
pe
of
talk
:
Design and Manufacturing in 2030
© Arm 20203
2030 will be “post Moore’s Law”
Moore’s Law ends: 2029
Yeric, 2015 IEDM(figure from The Economist)
© Arm 20204
5© Arm 2020
Motorola 68030© Arm 20206
© Arm 20207
© Arm 20208
2.5 minutes of video = 2 AA batteries
1 billion hourswatched per day
YouTube: 500 hours uploaded every minute
2020 spirit animal
© Arm 20209
There are now >500 hyperscale data centers
Home security camera footage alone will require 100 more near term
(yes, most of the video triggers will be cats)
© Arm 202010
>50% connectionsmachine-to-machine
Huawei to ship 600,000 5G AAUs5G: 1M devices / km2
26B+ IoT connections
© Arm 202011
Data poised to dominate
2015:Fixed Access networks: 167 TWhWireless networks: 50 TWh
https://www.cfr.org/blog/what-5g-means-energy
https://www.nature.com/articles/d41586-018-06610-y
2018: Data centers 200 TWh
© Arm 202012
$10,000
$100,000
$1,000,000
$10,000,000
$100,000,000
$1,000,000,000
1960 1970 1980 1990 2000 2010 2020 2030 2040
Moore’spaper
Power
Delay
Cost
© Arm 202013
$10,000
$100,000
$1,000,000
$10,000,000
$100,000,000
$1,000,000,000
1960 1970 1980 1990 2000 2010 2020 2030 2040
Moore’spaper
Cost
© Arm 202014
$10,000
$100,000
$1,000,000
$10,000,000
$100,000,000
$1,000,000,000
1960 1970 1980 1990 2000 2010 2020 2030 2040
Moore’spaper
© Arm 202015
$10,000
$100,000
$1,000,000
$10,000,000
$100,000,000
$1,000,000,000
1960 1970 1980 1990 2000 2010 2020 2030 2040
Patterning tool cost
contact
G- and I-line
ArF dry
ArF immersionKrF
EUV: Extreme Ultra-Violet
© Arm 202016
$10,000
$100,000
$1,000,000
$10,000,000
$100,000,000
$1,000,000,000
1960 1970 1980 1990 2000 2010 2020 2030 2040
Moore’spaper
contact
KrF
ArF dry
ArF immersion
EUV
High-NA EUV
G- and I-line
See: Moore’s 2nd Law, a.k.a. Rock’s Law© Arm 202017
High-NA EUV: a bargain at $255M
Current (smaller) EUV toolsRequire three Boeing 737’s to ship
© Arm 202018
High-NA EUV: a bargain at $255M
Current (smaller) EUV toolsRequire three Boeing 737’s to ship
© Arm 202019
What could come after high-NA EUV?
EUV information density: more than30 million UHD TV screensper wafer
1011
01001
111101
010101
© Arm 202020
What could come after high-NA EUV?
• 6nm EUV 2.0
• e-beam writing
11010
01011
00011
010110
• Nano-imprint
• Self-Assembly
© Arm 202021
DNA self-assembly
Eric Winfree Nadrian Seeman
Paul RothemundBindinglocation
+ =
© Arm 202022
3nm transistors
“DNA Origami” tiles
© Arm 202023
~23nm transistors
“DNA Origami” tiles
3-5nm precision placement
Fractal assembly of DNA origami
Lulu Qian
© Arm 202025
Fractal tiling of DNA origami
$255M
GrigoryTikhomirov $20
PhilipPetersen
20 nano-cents
http://qianlab.caltech.edu/FracTileCompiler/© Arm 202026
Fractal tiling of DNA origami
64 cores at 7nm:
NEOVERSE N1
20
mm
20mm2mm x 2mm
There’s plenty of room at the bottom! (1959)
DNA origami doesn’t need to be 2D
DNA bricks with programmable 3D pixels (voxels):
2.5 x 2.5 x 2.7 nmHarvard
Self-assembling dodecahedron (90% yield)TU Munich
© Arm 202028
130mm x 130mm x 130mm
2012: DNN wins ImageNet competition
https://www.nature.com/articles/d41586-018-06610-y
Alex Krizhevsky
© Arm 202029
2015: ResNet more accurate than people
https://www.nature.com/articles/d41586-018-06610-y
Human error ~ 4%
ResNet training: 10,000 PF29 hours using 8 Tesla P100 GPUs(300W each, so 70kWh total)
© Arm 202030
With $25M hardware, you can play 4.9 million games in 3 days
5000 TPU v1’s40W each
feeding64 TPU v2’s124W each
40 days
= 216 MWh
© Arm 202031
For the equivalent energy of training AlphaGo:
Cambridge to Lisbon to Rome to Warsaw back to Cambridge
Tesla S P85
(then 146 more loops)
(or power 23 homes for a year)
(or fly for 1000 hours)
© Arm 202032
https://openai.com/blog/ai-and-compute/
This is 7x/year
(and data are actually beyond linear)
If this were to continue,a quantum computerin 2029 would becompeting with2.8e11 PFs-day
280 Yottaflops-day
© Arm 202033
Progress in neuromorphic computingbrain inspired
Intel
Intel
© Arm 202034
• Epfl blue brain rpject
The “connectome”
Mouse Brain ~70m neurons
But number of synapses in cerebral cortex 60 trillion
100 - 100,000 inputs per neuron
The Blue Brain Project (BBP/EPFL)
© Arm 202035
© Arm 202036
DNA origami doesn’t need us
Energy and Delay scaling
© Arm 202038
39
IEDM 2017
© Arm 2020
HS Wong, et al., SISPAD 2009
Koomey’s LawRevision (2016)
© Arm 202040
https://semiengineering.com/whats-after-finfets/
Jacky Huangcoventor.com
© Arm 202041
Multi-layer transistor roadmap
© Arm 202042
© Arm 202043
Whatever happened to graphene?
Andre Geim Konstantin Novoselov
+ =2010Nobel Prize
2004:
Band Gap 0V: not a good semiconductor
© Arm 202044
51 new 2D semiconductors from one paper
2018
Band Gap
© Arm 202045
57 new quaternary* 2D semiconductors
2018
*quaternary:4 different elements
© Arm 202046
The wave of computational materials
• DFT: Density Functional Theory (Walter Kohn)
• Computational methods in quantum chemistry (John Pople)
1998 Nobel Prize
© Arm 202047
The wave of computational materials
3 days simulation per point
© Arm 202048
The –ic’s of 2030?
Valleytronics
Transmonics
Skyrmionics
Photonics
Spintronics
Twistonics
© Arm 202049
Superconducting electronics (SCE)
Flip-Flop: just 5 JJ’s(up to 750GHz)
1mV
© Arm 202050
https://www.iarpa.gov/index.php/research-programs/supertools
© Arm 202051
Disruptive technology toward 2030
Valleytronics
Transmonics
Skyrmionics
Photonics
Spintronics
Twistonics
© Arm 202052
From Lab to Fab
GAO-14-406SP (2014)
Lab to Fab →
© Arm 202053
“Research is critical to advancing semiconductor innovation in the U.S. American semiconductor design and manufacturing companies invest approximately one-fifth of revenue in R&D, almost $40 billion in 2019, representing the second-highest rate of research investment of any industry.”
© Arm 202054
Example from Arm Research
© Arm 202055
Example from Arm Research
Algorithms
Architecture
Microarchitecture
Materials
Circuits
Process integration
Device
Software
Application
“materials to systems”“atoms to applications”
© Arm 202056
Example from Arm Research: CeRAM100nm
© Arm 202057
DARPA’s Electronics Resurgence Initiative
https://www.darpa.mil/work-with-us/electronics-resurgence-initiative
© Arm 202058
Multi-level cell (MLC) memory
Vg(SET)=0.7V
Vg(SET)=0.75V
Vg(SET)=0.8V
Vg(SET)=0.85V
Vg(SET)=0.9V
Vg(SET)=0.95V
Vg(SET)=1.0V
Typical VREAD
• 300mm wafers• Sub-100nm dots• Low Power• Speed: < 10nS
© Arm 202059
Correlated Electrons
A true quantum phase transition is adiabatic
60K 1.8K
© Arm 202060
Nevill Mott1954-1971
New physics, new opportunities
© Arm 202061
NVM MLC example
Lab to Fab →
MLC is all about variation control• Best deposition/cost tradeoff?• PVD, CVD, ALD …• How do you etch it?
© Arm 202062
© Arm 202063