Big Data: Thoughts from the perspective of the semiconductor industry

Big Data SSGJanuary 24, 2012

Big Data: Thoughts from the perspective of the semiconductor industry

Celia Merzbacher, SRC VP for Innovative Partnerships

2

Happy Data Innovation Day!

3

4

5

80Gb cost $9,000,000 !!!

in 1982 dollars

1982: Best available storage technology was the IBM 3350

iPod(5G) 80GB

2012

126 IBM 3350’s = storage in

1 iPod

80Gb cost <$100

in 2012 dollars

Each unit: 635 MB $70,000

Hardware Advances Enable Big Data

Moore’s Law: # transistors/chip doubles every 24 months

6

In 1982…Aug 17—the first compact disc goes on sale

7

Oct 1—Sony launches the first compact disc player

In 1982…Time magazine’s Man of the Year was…

8

The Computer

9

But in 1982 the U.S. semiconductor industry saw threats on the horizon.

US semiconductor market share was dropping…Federal funding for academic research on silicon was declining…

The pipeline of talent was drying up.

Federal $

Research

Talent

Market Share

Semiconductor Market Share

0 %

20 %

40 %

60 %

80 %

US

Japan

In the early 1980’s

In 1982…

11Robert Noyce Jack KilbyErich

Bloch

The Semiconductor Research Corporation (SRC) was launched with the support of visionary industry leaders.

Objectives:Define relevant research directions Explore potentially important new technologies Generate a pool of experienced faculty &

relevantly educated students STAY ON MOORE’S LAW

Moore’s Law: 1971-2011

1970 1975 1980 1985 1990 1995 2000 2005 2010 20151,000

10,000

100,000

1,000,000

10,000,000

100,000,000

1,000,000,000

10,000,000,000

R² = 0.94986565616323

Year

MPU

tran

sisto

rs

Cu in

terc

onne

cts

High

-K g

ate

insu

l.Pb

-free

pac

kagi

ng

FinF

ET

Triple Core

Dual Core

Quad CoreHex Core

Eight Core

2ln 2 2.13

0.3252Y years

1E+08 1E+15 1E+221E+04

1E+10

R² = NaN

b, bit/s

m, IPS

(Inst

ruct

ions

per

sec

-on

d)

Benchmark capability m (IPS) as a function of b (bit/s)

trN fb

Power is the main issue for further scaling of high-performance computing

~100 W

Scaling up performance

IBM Watson supercomputer The most recent and most impressive

demonstration of an artificial intelligence computer system Capable of answering questions posed in natural

language Winner of the Jeopardy! quiz show in 2011

~3000 processor cores (POWER7) each consisting of 1.2B transistors and operating at

3.5GHz approximate total binary throughput ~ 1022 bit/s.

14

& power

~200kW of power EPA estimated data centers used 1.5% of total

US electricity in 2007

1E+08 1E+15 1E+221E+04

1E+10

R² = NaN

b, bit/s

m, IPS

(Inst

ruct

ions

per

sec

-on

d)

Benchmark capability m (IPS) as a function of b (bit/s)

1014 IPS1019 bit/s30 W

Basic algorithms need to work in very few steps! (L.G Valiant, A quantitative theory of neural computation, Biol. Cybern. (2006) 95

~100 W

Estimates of computational power of human brain:Binary information throughput:

b ~1019 bit/sGitt W, “information - the 3rd fundamental quantity”, Siemens Review 56 (6): 36-41 1989(Estimate made from the analysis of the control function of brain: language, deliberate movements, information-controlled functions of the organs, hormone system etc.

Number of instruction per second

m ~ 108 MIPSH. Moravec, “When will computer hardware match the human brain?” J. Evolution and Technol. 1998. Vol. 1 (Estimate made from the analysis brain image processing)

How can we decrease the energy needed to move/store data?What can we learn about information processing from Nature?

1000x algorithmic efficiency

Infrastructuralcore

The IT Platform of Today:Mobiles at the Edge of the Cloud

[J. Rabaey, ASPDAC’08]

MobileAccess

The Cloud

Mobile data growth[Source: Cisco VNI Mobile, 2011]

Mobile traffic grew 2.6x in 2010 (nearly tripling for 3rd year)Driven by Tablets

The SwarmInfrastructuralcore

The Swarm at The Edge of the Cloud

[J. Rabaey, ASPDAC’08]

MobileAccess & Relay

The Cloud

New STARnet* Center: TerraSwarmDirector: Ed Lee, UC-Berkeley

*STARnet is a subsidiary of SRCwww.src.org/program/starnet/tsrc/

Si-mCell: A hypothetical 1-mm3 Si computer

Memory: 40 kbitLogic: 320 bit

1mm

Logic MemoryFmin (nm) 4.5 10

N 320 40,000Ebit (J/bit) 3×10-18 ~10-15 (read)

Ecycle (J) ~10-15 ~10-13 (read)fclk, MHz 100Pactive (W) 10-7 10-5

Pleak (W) 6.4×10-7 assumed lowPtotal (W) ~1.1×10-5

Qactive (W/cm2) 2 170

Qleak (W/cm2) 11 assumed lowQtotal (W/cm2) ~200

Memory access is the most severe limiting factor of Si-mCell due to line charging

Exceeds capability of known cooling techniques

20

Nanoelectronics Research InitiativeFinding the Next Switch

UC Los AngelesUC BerkeleyUC IrvineUC RiversideUC Santa Barbara

Notre Dame PurduePenn State UT-Dallas

UT-Austin RiceUT-Dallas NCSUU. Maryland Texas A&MGIT

SUNY-Albany

Purdue U. VirginiaHarvard GITColumbia MIT

Over 40 Universities in 19 States

(co-funds all centers)

Virginia Nanoelectronics Center (ViNC) University of Virginia Old Dominion University College of William & Mary

BrownU. AlabamaNorthwestern

Columbia Carnegie Mellon Illinois-UC MITStanford Notre Dame (2)Nebraska-Lincoln Columbia / U. FloridaPenn State U. of MinnesotaPrinceton / UT-Austin Cornell / PrincetonUC-Santa Barbara Drexel University / UI-UC / U. PennUC-Riverside / GeorgiaVirginia Commonwealth / UC-R / Michigan / U. Virginia UC-Riverside / UC-I / UC-SD / Rochester / SUNY-BuffaloU. Pittsburgh / U. Wisconsin-Madison / Northwestern

NRI Nanoelectronic Devices

HeterojunctionsNotre Dame, Penn State

NanowiresPenn State

GrapheneNotre Dame

Tunnel DevicesMIND

Spin-Wave DeviceWIN - UCLA, UCSB

Spin-FETWIN - UCLA

Nanomagnet LogicMIND - Notre Dame

WIN - Berkeley

Spin-Torque DeviceWIN - UCI

PtMn

Co70Fe30

RuCo40Fe40B20

MgOCo60Fe20B20

12oH

+Idc

-100 0 100 2003

4

5

6

7

8

R (

k)

Direct current ( mA)

T = 100 KH = 1.4 kOe

0 1 2 3 4 5-0.015

-0.010

-0.005

0.000

T = 100 KH = 1.4 kOeIac = 5 mAIdc= 0 mA

V dc (m

V)

Frequency (GHz)0 1 2 3 4 5

0

2

4

6

8T = 100 KH = 1.4 kOeIac = 5 mAIdc = 100 mA

V dc (m

V)

Frequency (GHz)

A B

C D

Bilayer pseudoSpinSWAN - UT Austin

All-Spin Logic INDEX - Purdue U.

Graphene PN Junction Device

INDEX - SUNY Albany

A

F

U = ‘1’

B C

(a)

y

x

z

‘0’U =

VGnVn

VGp

Vp

Graphene Integration

INDEX – SUNY Albany

Graphene

Tunneling Insulator

Graphene

FMD

Graphene

Substrate

Contacts Insulators Oxidation

Graphene Processes SWAN – UT Dallas

Device and Architecture Benchmarking MIND/WIN/INDEX/SWAN – Led by K. Bernstein, IBM 21

Breakthrough Technology Challenges for next decades

From fundamental physics it seems likely that the scaling of current technology will end in the few nanometer regime. NRI is working to develop replacement technologies So far, a replacement technology has not been found

Are there other models for information processing technologies that offer the promise to sustain Moore’s Law?

Looking to organic systems, i.e., at the intersection of chemistry, biology, and information processing

ACBSSpecifications of a Human Cell

- 10 mm overall size- - 0.36 nm between base pairs in DNA. Average protein is 5 nm.

- 107 biochemical operations per second- 1 pWatts power consumption- 30,000 node gene-protein molecular network with nanoscale

devices.- 20 kT per molecular operation

(vs. 104–105 kT in advanced nanoelectronics)- Functions: sensing, communication, actuation, feedback

regulation, molecular synthesis & transport, detoxification, defense, self assembles from a single embryonic cell.

Biology computes efficiently and precisely with noisy and unreliable components on noisy real-world signals.

Courtesy of Rahul Sarpeshkar, Analog Circuits and Biological Systems Group, MIT

Memory

Nature Has Been Processing Information for a Billion Years

Logic

L

L

L

L

L

L

L

L

LL

LL

L

M (DNA)

S

L

L

L

L

L

E

E

E

E

C C

E E

Si-mCell

V=1mm3

Bio-mCell – A Living Cell

Our studies show that the Si-mCell cannot match the Bio-µCell in the density of memory and logic elements, nor operational speed, nor operational energy:

Memory: 1000x moreLogic: >10x morePower: 1,000,000x lessAlgorithmic efficiency: 1000x more

About 500 of these cells would fit in the cross-section of a human hair

25

DNA-inspired memory DNA volumetric memory density far exceeds (1000x) projected

ultimate electronic memory densities Potential for very low-energy memory access Goal: Demonstrate a miniaturized, on-chip integrated DNA

storage

HardDiskDrive NAND flash DRAM DNA in cellRead/Write latency 3-5 ms/bit ~100ms/bit <10 ns/bit <100ms/bitEndurance (cycles) unlimited 104-105 unlimited unlimited

Retention >10 years ~10 years 64 ms >10 yearsON power (W/GB) ~0.04 ~0.01-0.04 0.4

Aerial Density ~ 1011 bit/cm2 ~ 1010 bit/cm2 ~ 109 bit/cm2 n/a Volumetric Density n/a 1016 bit/cm3 ~1013 bit/cm3

<10-11

1019 bit/cm3

26

DNA: The Ultimate Hard Drive?

http://www.wired.com/wiredscience/2012/08/dna-data-storage/

DNA Memory

Researchers stored an entire genetics textbook in less than a picogram of DNA — one trillionth of a gram — an advance that could revolutionize our ability to save data.

5.27×106 bit

DNA memory can be stable 100+ years

European Bioinformatics Institute, Wellcome Trust Genome Campus, Hinxton CB10 1SD, UKAgilent Technologies, Genomics–LSSU, 5301 Stevens Creek Boulevard, Santa Clara, California 95051, USA

Encoded into DNA code computer files totaling 739 kilobytes of hard-disk storage and with an estimated 5.2 × 106 bits

Synthesized and sequenced the DNA, and reconstructed the original files with 100% accuracy.

Storage scheme is theoretically scalable beyond current global information volumes

Current trends in DNA synthesis costs should make the scheme cost-effective for sub-50-year archiving within a decade.

Possible New SRC Initiative: SemiSynBio

28

Organizing Committee: Rahul Sarpeshkar/MITTimothy Lu/MITSami Issa / ATICAndrew Hessel / AutodeskEric Klavins / U. WashingtonLarry Sumney / SRCSteven Hillenius / SRCRalph Cavin / SRCVictor Zhirnov / SRC

Exploring potential benefits to the semiconductor industry arising from synthetic biology

Meeting Date: February 22&23, 2013Meeting Place: Cambridge, MA

Take Away Messages

29

Success of Big Data will depend on continued advances in computation hardware, aka semiconductors

Moore’s Law (for CMOS) is facing physical limits Power is the main issue for further scaling of high-

performance computing There are no evident replacement technologies

Nanoelectronics research is seeking new devices New research turns to biology

DNA-based memory Using/mimicking Nature in other areas may allow

Moore’s Law (for performance) to continue “beyond CMOS”.

Industry—through SRC—continues to fund leading edge university research in partnership with Government

30

SRC Creates Value Through Partnerships

• Maximizes technological progress• Leverages investments• Utilizes the strengths of each sector • Expands and replenishes the

professional community

Tactical Perspective, “Can-Do” Attitude,

FUNDING

Industry

Strategic Perspective,National Needs,

Credibility, FUNDING

Government Universities

Creativity, Faculty Expertise, Student Resources

Semiconductor Research Corporation: A Family of Distinct, Related Program Entities

Updated January 2013

Each entity has a distinct set of member companies and Government partners.For more information go to www.src.org

Global Research

Collaboration

Ensuring vitality of current industry

Focus Center Research Program Phase VI

STARnet

Early research engagement of

key long horizon

semiconductor challenges

Energy Research Initiative

Emphasis on efficient/clean

energy generation, storage and distribution

Education Alliance

Attracting and educating the

next generation of innovators

and technology leaders

Nanoelectronics Research Initiative

Beyond CMOS –the next switch and associated architectures

32

Benefits of SRC Approach (to Univ & Govt)

Value of research is enhanced Provide insight on industry needs to researcher community Input and feedback from industry at periodic reviews E-seminars and e-workshops facilitate near real-time sharing of

research results and tech transfer Interactions and opportunities for personnel exchanges among

universities and industryStudent education is enhanced Industry liaisons & mentors engage with students Participation in TECHCON, SRC’s annual technical conference at

which students present research and network with industry representatives.

Opportunities for student internships at SRC member companies.

Explore new research directions E.g., joint workshops, new program “spin offs”, etc.

SRC created an industry-guided global university research ecosystem

Since 1982… Over $1.6B invested by SRC

participants 9,195 students 2,025 faculty members 261 universities in 27 countries

1500 students 500 faculty 120 universities worldwide

In 2012…20X

increase over 1982

34

Current SRC Member Companies

Backup

35

Essential SRC Features Industry-driven, consensus-based goals embodied in:

• Moore’s Law• ITRS (International Technology Roadmap for Semiconductors)

Focus on pre-competitive university research (>5 yr time horizon) Members have rights to resulting IP Involves the current industry experts (provide input/ feedback/

oversight and tech transfer) Managed by an independent entity (facilitates interactions among

members and with universities & government agencies) Nimble and adaptable (~1/3 of projects turn over annually) Accountable; value-driven; efficient; effective Attracts world-class researchers (faculty & students)

36