Transforming Concepts To Reality: New Technologies for New

Max M. Shulaker

Transforming Concepts To Reality:

New Technologies for New Applications

Massachusetts Institute of Technology

Data ExplosionE

xa

B(B

illio

ns o

f G

B)

0 4

0K

2006 Year 2020

Raw unstructured dataWide variety & complexity

“Swimming in sensors, drowning in data”

2

• Mine, search, analyze data in real-time

Data centers, mobile phones, robots

Many Obstacles Simultaneously

Power Wall Memory Wall

Execution time

Compute

Memory access

10M

Hz

5G

Hz

3

Also:

interconnect wall,

complexity wall,

resilience wall…

Communication Wall

1980 2010

4

Enabling New Applications:

new sensors

better transistors

novel memories

new architectures

thermal management

improved algorithms

yield and reliability

Not

Not

Not

Not

Not

Not

NotNanosystem

s

5

Solution: NanosystemsTransform new nanotech

into new systemsenabling new applicationsNew Devices

New Fabrication

New Sensors

Abundant-Data

Applications

Revolutionary

Architectures

a

6

Systems Today

2-Dimensional

memorysensor

computing

7

Future Nanosystems

8

Storage (memory)

Increased Functionality

Fine-grained

3D integration

Computing Logic

Impossible with today’s technologies

Future Nanosystems

9

Storage (memory)

Increased Functionality

Fine-grained

3D integration

Computing Logic

Realizing Nanosystems TODAYEnabled by Emerging Nanotechnologies

10

~1.2nm

CNT

Gate-oxide

• Energy efficiency: ~10X benefit [1]

• Chemical sensors: ideal candidate

1 [Chang IEDM 2012]

2 µm

Gate

2 µm

Gate

Carbon Nanotube FET (CNFET)

Ideal CNFET Inverter

P+ Doped

N+ Doped

INPUT

11

Ideal CNFET sensor

Functionalization(metal porphyrin, DNA,

SAMs, polymers, etc.)

INPUT

12

BIG Promise, BUT Major Obstacles

Mis-positioned CNTs Metallic CNTs

13

Combined processing + design solutions

Most Importantly

• VLSI processing

No per-unit customization

• VLSI design flow

Immune CNT library

14

[Shulaker Nature 13] Collaborators: G. Hills, N. Patil, H. Wei, H. Chen, H.-S.P. Wong, S. Mitra

CNT Computer

15

Concurrent programs: Sort, Count

CLK1

A[0]

A[1]

A[2]

B[0]

B[1]

B[2]

CLK1

A

B

B-A

[0]

[1]

[2]

[3]

[4]

Ne

xt

Instr

Ad

dr

Data fetch addresses ALU result & next instruction

bit

bit

MIPS

instructions

emulated

• AND• ANDI

• BLEZ• BLTZ

• J• LB

• OR• ORI

• SLL• SLLV • SRLV

• SRL • XOR• XORI

• BGEZ • BNE • NOOP • SB • SRA • SUBU

Multi-Tasking

16

17

CNFET Gas SensorsR

ela

tive

sen

sit

ivit

y

Functionalization:

How do we do better?

18

• Massive ILV density > TSV density

Monolithic

Nano-scale

inter-layer vias (ILVs)

TSV (chip stacking)

Through silicon via

(TSV)

3D Integration

19

logic

memory

ILVs

Monolithic 3D: Logic + Memory + Sensing

• Increased data bandwidth

20

• Memory access time

Less processor idle time

• Memory access energy

Memory + logic stacking

• Resource Contention

Wide connectivity

Energy and Performance Benefits

21

• Low-temperature fabrication: <400 °C

Major obstacle for silicon CMOS

Realizing Monolithic 3D

22

Logic/ Sensing Memory

CNFETs RRAM


23

<200 °C <200 °C

• Low-temperature fabrication: <400 °C

+ +Emerging

Logic/ Sensors

Emerging

Memory

Monolithic 3D

Integration


24

Naturally enabled

• Combine device + architectural benefits

25

3D NanoSystem

memory

sensing

logic

logic

• First monolithic 3D system

• >2 Million CNFETs, 1 Mbit RRAM

[Shulaker Nature 17]

26[Shulaker Nature 17]

Wafer-scale design + fabrication

1 2 3

4 5 6

7 8

NOT Just a Cartoon

27

3D NanoSystem: Fabrication

• Starting substrate

28

• Silicon FETs

1050 ºC dopant activation


1. Bottom layer only

2. Silicon OR CNFET

29

• Inter-layer dielectric (ILD)


30

• Inter-layer vias (ILVs)

Ultra-dense vertical inter-connects


31

• CNFET logic


Memory access circuitry

Classification acceleratoron-chip computation

32

• 2nd ILD + ILVs


• Memory: RRAM

Pt/HfOX/TiN

33


RRAM array: 1 Mbit

34

• 3rd ILD + ILVs


35

• CNFET logic


CNTs X100,000

36

• CNFET functionalization


>1 million gas sensors

37

Integrated Computation + Memory + Sensing

CNT computing logic

Memory

Ultra-dense

vertical connections

Millions of sensors

Terabytes / second

Abundant sensor data:

Extensive + accurate classification

38

Experimental Demo

Vacuum

chamberElectrical

feedthrough

to chip

Bubblers

Gas Flow Control


39

Raw Data

Lemon Juice

Rubbing Alcohol

Sensor data written

to RRAM array

0 64 1280

50

0 64 1280

50

Measured Output

clock cycle

Fe

atu

re V

alu

eF

ea

ture

Va

lue

CNFET classification accelerator


Feature 2

Feature 1

40

Experimental Results

Nitrogen

Lemon juice

Vinegar

Rubbing Alcohol

Vodka

Wine

Beer-0.5 0 0.5

-0.5

0

0.5

feature 1 (normalized units)

fea

ture

2 (

no

rma

lize

d u

nits)

PCA classificaiton

gas 1: trial 3

gas 2: trial 1

gas 2: trial 2

gas 2: trial 3

gas 3: trial 1

gas 3: trial 2

gas 3: trial 3

gas 4: trial 1

gas 4: trial 2

gas 4: trial 3

gas 5: trial 1

gas 5: trial 2

gas 5: trial 3

gas 6: trial 1

gas 6: trial 2

gas 6: trial 3

gas 7: trial 1

gas 7: trial 2

gas 7: trial 3

Principle Component Analysis (PCA)


3D NanoSystem: Take-Away

41


42

• Energy-efficient logic + memory

Heterogeneous integration


43



• High bandwidth communication

Fine-grained 3D integration




• High bandwidth communication

Fine-grained 3D integration

• Transform massive data into useful info

Sensing immersed in computation + memory

44

Massively Parallel Sensing Immersed in Computation

MIT-ADI-MGH Collaboration

Memory

Millions of sensors

Computing

Health

Monitoring

Cancer

Screening

Early Disease

Detection

Infectious Disease

Tracking

Big-Data Healthcare Analytics:

New Technologies New Applications

46

Conclusion

• Nanosystems useful today

Exciting opportunities

• New solutions: elegantly simple

Combined processing + circuit design

47

Future Nanosystem

Prototypes

First Nanosystem

Demos

CNTs

High Performance

CNFETs

Documents

Transforming Concepts To Reality: New Technologies for New