abundant-data computing the n3xt 1,000x
TRANSCRIPT
Abundant-Data Computing
The N3XT 1,000X
Department of EE & Department of CS Stanford University
Subhasish Mitra
Solution: NanoSystems
2
Transform new nanotech into new systems enable new applications New devices
New fabrication
New sensors
imperfections? large-scale fabrication? variability?
New Architectures
a
Abundant-Data Explosion
Exa
B (B
illio
ns o
f GB
)
0
40
K
2006 Year 2020
Unstructured data Wide variety & complexity
“Swimming in sensors, drowning in data”
● Mine, search, analyze data in near real-time
▪ Data centers, mobile phones, robots 3
Abundant-Data Applications
4
Memory wall: processors, accelerators
Compute Memory
ResNet-152 (CNN)
Deep Learning Accelerators AlexNet (CNN)
…20%
80%
15%
85%
8%
92%
Language Model (LSTM)
5
Nano-Engineered Computing Systems Technology
[Aly IEEE Computer 15]
Computation immersed in memory
6
N3XT NanoSystems
Memory
Increased functionality
Fine-grained, ultra-dense 3D
Computing logic
Impossible with today’s technologies
Carbon Nanotube FET (CNFET)
7
CNT: d = 1.2nm
2 µm
Gate
2 µm
Gated
CNFET
Sub-litho pitch
2. First CNT computer
[Shulaker Nature 13, ISSCC 13, IEDM 14]
1. Energy Delay Product
~10X benefit Full-chip case studies
[IBM, IMEC, Stanford, others]
Example: OpenSPARC T2
8 [Stanford + IMEC, Unpublished]
0.05
0.5
0.1 1 10
tota
l ene
rgy
per c
ycle
(n
J)
clock frequency (GHz)
FinFET Nanowire FET
CNFET
preferred
Putting into Perspective
9
• Existing technology benchmarking + CNFETs
4.E-1
4.E+0
4.E+1
1.E+2 1.E+3 1.E+4
Si-CMOS high performance
“beyond” CMOS
Si-CMOS low power
40
4
0.1 10 1 0.4
adder frequency (GHz) adde
r ene
rgy
per o
pera
tion
(fJ)
preferred
vdWFET
ExFET HetJTFET
gnrTFET GaNTFET
BisFET
TMDTFET
PiezoFET
NCFET
ITFET GpnJ
ThinTFET HomJTFET
32-bit adder [Nikonov & Young, 2013 & 2015]
3D Integration
10
l Massive ILV density >> TSV density
Nano-scale inter-layer vias (ILVs)
TSV (chip stacking)
Through silicon via (TSV)
Dense, e.g., monolithic
Device + Architecture Benefits
11
Top Electrode
Metal Oxide
Btm Electrode + + Emerging
logic Emerging memory
Monolithic 3D integration
Naturally enabled
[Wei IEDM 13, Shulaker VLSI Tech 14]
l Low-temperature fabrication: < 400 °C
In-situ classification: Extensive, accurate
First 3D NanoSystem
[Shulaker Nature 17] 12
CNFET compute accelerator
(classification)
Millions of sensors 1 Mbit
RRAM
Abundant data: Terabytes / second
No TSVs
>2 Million CNFETs, 1 Mbit RRAM
N3XT Simulation Framework
13
Joint technology, design & app. exploration
Architecture exploration
Energy, exec. time
Thermal
Physical design, yield, reliability
Heterogeneous technologies
System-level analysis Abundant-
data apps
14
~1,000X benefits, existing software
chip stacking: 2-4x benefits
IBM graph analy/cs
1×
10×
100×
PageRank Connected Components
Breadth-‐ First Search
Linear Regression
Language model
(LSTM Neural Network)
AlexNet (Neural Network)
Energy Execu+on Time
Bene
fits
851× 400× 510× 970× 1,950× 210×
Massive Benefits: Deep Learning, Graph Analytics, …
More Opportunities
15
Accelerators
Brain-inspired
Technology innovations
“Brain-Inspired Computing Exploiting Carbon Nanotube FETs and Resistive RAM: Hyperdimensional Computing Case Study,” ISSCC 2018.
Conclusion
16
l Nanosystems today
l Game ON, to the era
l N3XT 1,000X
§ Compute + memory + sensing
§ Densely interwoven