optical semiconductor devices to Year 2025to Year …If anybody likes to keep this talk’s slides, please email me!Roadmap of ultrafast energy-saving optical semiconductor devicesto

If anybody likes to keep this talk’s slides, please email me!

Roadmap of

ultrafast energy-saving optical semiconductor devices

to Year 2025to Year 2025

--- <speed, energy, size> estimates of optical Micro Processor Unit ---

Yoshiyasu UENO

National Univ. of Electro-Communications (UEC)

Tokyo, Japan

42 d d T k S t 22 2010

Ultrafast Optical Logic Lab., UECUltrafast Optical Logic Lab., UEC1

42nd ssdm, Tokyo, Sept. 22, 2010

Contents

[1] Total energy consumptions in ICT technology

Contents

[ ] gy p gy

[2] Recent trends of optical-data-processing devices <speed, energy, size>

[3] All-Optical gates: from principles to new potentials <speed, energy, size>

[4] Crude estimates about “optical MPU” (long-term research)

Co authors and Collaborations

(first time, too, to our knowledge)

Co-authors and Collaborations


[1] Energy consumptions in ICT-related systems

[1] Impacts of ICT energy consumptions

[1] Energy consumptions in ICT-related systems

Conventional Technol.

S

Around 2005,Clk freq. reached 3 GHz.

(electronic MPU’s)

In Data Centers of Google, Microsoft, etc.

Nuclear Power-Station1 GW / reactor

D l/

Nano-materials inside MPU’s are nearly melt.（37 years of Moore’s law, IEEE 2008)

ata Ce te s o Goog e, c oso t, etcClk freq.: ３ GHz, with similar MPU’s

Heat-energy dissipation: 10 kW / sever rackData-center energy: 20 MW / data center3.0

4.0

peed

(GH

z)

Si l

Dual/Quadsaturating

speed

0.0

1.0

2.0

rose

ssor

clk

sSingle processor

Cost of DRAM

We need ultrafast & energy-savingOptical Transistors

3

Pr200819901980 2000

Calendar Year

pin future !

Ultrafast Optical Logic Lab., UECUltrafast Optical Logic Lab., UEC

Electric-energy consumptions, 1970-2006[1] Impacts of ICT energy consumptions

1000.0

(EJ)

y

100.0

he c

ount

ry (

elec

tric

ity

10.0

系列1

系列2

系列3

gene

ratin

gsu

pply

in th Japan

P.R. China

IndiaJapanply,

for e

1 0

系列4

系列5

系列6

Sup

ply

for

ele

ctric

ity s EU (15)

UK

USA

W ld

UK

ergy

sup

p

1.0 系列7

ary

Ene

rgy

annu

al to

tal World

mar

y en

e

0.1

1970 1980 1990 2000 2010

Prim

ath

e a

Year*1) Source: International Energy Agency (IEA), Paris,

Yoshiyasu Ueno, February 2010

Prim

1970 20101990


) gy g y ( )Energy balances of OECD countries and non-OECD countries.

*2) 1 EJ= 1×1018 J.

Macro-scopic:Primary Energy Supplies (sum of electricity and non-electricity), 2006


90

100

2006

)

Primary Energy SuppliesRatios of energy for electricityChina

USA

EU

50

60

70

80

90

rgy

Sup

ply

(EJ,

2

90

100

EJ,

200

6) for non-electricity

for electricityincl.

JapanChina EU

China

USA

0

10

20

30

40

tal P

rimar

y E

ner

50

60

70

80

rgy

Sup

ply

(

Vehicles,Jets, etc.

of today

Japan

ChinaEU

Japan China India EU15 UK USATot

20

30

40

mar

y E

ner

12,000

06)

GDP’s (2006)Electric Vehicles,next ??Chi

USAEU

Japan

India

39%

43%

36%39%

0

10

Japan China India EU15 UK USATota

l Pri

*1) Source: International Energy Agency (IEA) Paris

6,000

8,000

10,000

×10

9U

SD, 2

00

next ??

JapanChina 39%

*1) Source: International Energy Agency (IEA), Paris,Energy balances of OECD countries and non-OECD countries.

*2) 1 EJ= 1×1018 J.

0

2,000

4,000

Japan China India EU15 UK USA

GD

P (×

Data Centers in USA consuming 1.5% of all electricity (D. Miller, Stanford).not very large?? 1 5% correspondes to 10 nuclear reactors!!


not very large?? 1.5% correspondes to 10 nuclear reactors!!We need energy-saving devices.

Micro-scopic:one origin of ICT-energy consumptions


Zetta1 E+20

1.E+21

Tr-number times Clk-freq. has evolvedby a factor of 106×104= 1010 (in 40 years)Tr・Clk Product

Tr number: 106

Exa1.E+17

1.E+18

1.E+19

1.E+20

• FLOPS speed has increased by 1010.

Tr number: 106

Clk freq.: 104Moore’s magic

Peta

NEC Earth simulator ->

Cray Jaguar,IBM Roadrunner ->

NEC-Sun Tsubame ->

1 E+13

1.E+14

1.E+15

1.E+16

• MIPS speed has increased by 104, only,without supported by Tr number. (reasonable)

Tera

1.E+10

1.E+11

1.E+12

1.E+13

Pentium Pro(1996)

Pen 4(2002)

Core 2 Quad(2007) • Already relying on parallel-processing MPU’s and software.

many-folded parallel-structures are probably pushing-upelectric-energy consumptions (and costs).[hard to quantitatively characterize now though ]Giga

Cray-1 ->

1.E+06

1.E+07

1.E+08

1.E+09

80386(1985)

(1996)

8086(1978)

[hard to quantitatively characterize now, though.]

? for 2010-2050: Increasing demands are

1.E+04

1.E+05

1.E 06

1960 1970 1980 1990 2000 2010 2020 no. of Trs. times clk freq. 10x every 4 yrs.FLOPS 10x every 4 yrs.

Y. Ueno, February 2010

? for 2010-2050: Increasing demands are

FLOPS-type demands, or, MIPS-type demands ??

6Ultrafast Optical Logic Lab., UECUltrafast Optical Logic Lab., UEC

y yInstructions per sec. 10x every 7 yrs.Sources: FLOPS from measured results, MIPS from wikipedia.

to move from electronics to optics: its famous weaknessLatest degrees of integrations for optical-processing devices


g g p p g

Number of devices on one chip= 200 (Infinera, 2006) evolving steadily, driven by industrial demands,g y, y ,

and approaching the number 2,300 in intel 4004 (1971).2,300

ents

Electronic processorIntel 4004 (1971)

l com

pone

ne c

hip)

200

# of

opt

ical

(in o

n

Eindhoven Univ. Technol., Netherlands

#

source:M K Smit (Eindhoven U Technol ) IEEE LEOS Annual 2008

CIP Technologies, UK.


M.K. Smit (Eindhoven U. Technol.), IEEE-LEOS Annual 2008.

[2] Optical data processing “gates and memories”[2] Optical-data-processing gates and memories

Y. Ueno, UEC, 2010

All-Optical circuits w/ gates & memories

Optical data Optical data

Optical data

Drive energy (electric dc-bias)


[2] Optical-data-processing “gates and memories” <speed, energy, size>

[2] Latest optical gates and memories

(2-1) Optical buffer memories (fundamental-research)

25Gb/s, 1pJ/bit, (30μm)2

M.T. Hill (Smit Gr., Eindhoven),40-100Gb/s, 3pJ/bit, (10μm)2

T Katayama (Kawaguchi Gr ) 2009

40-160Gb/s, L= few mm longE. Kehayas (Dorren Gr., Eindhoven)

planar, ring-laser.T. Katayama (Kawaguchi Gr.), 2009,

Vertical, VCSEL. planar, coupled-gates.

photonic-crystal gate(FESTA, U. Tsukuba, AIST)

photon-electron interactionsIn bulk or MQW semiconductors

are used.


K. Asakawa et al., J. of Phys. 2006



(2-2) All-optical gates (for practical signal-conversion, 2R/3R, demux, etc.)

SMZ-DISC scheme, with non-linear cross-phase modulation XPM

original-SMZ scheme, with XPMGate, 2000年

input output

0.2

0.3

0.4

0.5

0.6

Sign

al (a

.u.)

0.2

0.4

0.6

Sign

al (a

.u.)

Gate, 2006年

Gate, 2009年

input output

input output

-50 0 +500.0

0.1

Delay (ps)-50 0 +50

0.0

Delay (ps)

168Gb/s, 2 pJ/bit, L≅ 1mmS. Nakamura, Ueno, Tajima (NEC),

input output

wavelength-conversion

640Gb/sT. Hirooka (Tohoku U. & NEC), Demux.

320Gb/s, 2.5 pJ/bit, L≅ 1mmY. Liu (Eindhoven), wavelength-conv.


T. Hirooka (Tohoku U. & NEC), Demux.



Converter XOR AND 2R/3R etc

(2-2) All-optical gate (fundamental-research in our univ. UEC, Tokyo)

SMZ-DISC scheme (XPM) in our group

Clock oscillator (UEC)

Flip-Flop (Eindhoven, Tsukuba)

Converter, XOR, AND, 2R/3R, etc.

-20 -10 0 +10 +20Distance in air, z (mm)

Wavelength (nm)


Gate, 2008年 Oscillator, 2005-2008年

1.0

1.5

y (a

. u.)

Gated waveform, 1540 nm1 0 0 1 0 0 0 1 0 1 1 1 …

-20

0 1546154815501552

Bm

/nm

)

g ( )

550 GHz

0.5

Inte

nsity

-60

-40

Inte

nsity

(dB 550 GHz

-75 -50 -25 0 +25 +50 +750

Time (ps)193.0 193.5 194.0

Optical frequency (THz)

200Gb/s, 3 pJ/bit, L≅ 1mm Δ2ps/40GHz optical clock oscillatorSuzuki Nakamoto et al (CLEO2006 NANO2008)

(160G-class)



Sakaguchi, et al. (Opt. Comm. 2009) Suzuki, Nakamoto, et al. (CLEO2006, NANO2008)

200-Gb/s gated waveforms,

in the middle of our experimental studies


in the middle of our experimental studiesJun Sakaguchi, et al., Opt. Comm. 2009.

good !!

b tt

Distance, z (mm)good !!

better(accelerated)

y (a

. u.)

badly gated(too slow) In

tens

ity

Time (ps)Data-pattern-induced

amplitude noise




COE-research: DISC-loop-type mode-locked pulse source2-ps, 40-GHz pulse and comb generation, 2005-2006

R. Suzuki, et al., CLEO, Long Beach, USA, May 2006.

.).) 00Auto-correlation trace Optical-frequency-comb spectrum

0.5

1.0

tens

ity (a

. u.

13.5 dB

25ps = 41GHz

0.5

1.0

tens

ity (a

. u.

13.5 dB

25ps = 41GHz

-60

-40

-20

nsity

(dB

m) 330pm = 41GHz

-60

-40

-20

nsity

(dB

m) 330pm = 41GHz

Original scheme from,Y. Ueno, et al., Appl. Phys. Lett. Oct. 2001 -37.5 -25 -12.5 0 +12.5 +25 +37.50.0

Delay (ps)

SHG

Int

-37.5 -25 -12.5 0 +12.5 +25 +37.50.0

Delay (ps)

SHG

Int

1546 1548 1550 1552 1554

-80

60

Wavelength (nm)

Inte

n

1546 1548 1550 1552 1554

-80

60

Wavelength (nm)

Inte

n

y (p )y (p )measured t-f product, Δt⋅Δf= 0.53

(1.2 times that of a Gaussian pulse)

20

0

m)

20

igna

l V

)

-50 -25 0 25 500

10

20

Delay (ps)Aut

o-co

rrel

atio

n Si

gnal

In

tens

ity (

mV

)

1553 1554 1555 1556 1557

-80

-60

-40

-20

0

Wavelength (nm)

Inte

nsity

(dB

m)

(3) Gloop> 0dB

0m]

Proof of principle (1), threshold behavior Proof of principle (2), linearly controlled pulse widths6ps

)

6ps)

1553 1554 1555 1556 1557

-80

-60

-40

-20

0

Wavelength (nm)

Inte

nsity

(dB

m)

-50 -25 0 25 500

10

20

Delay (ps)Aut

o-co

rrel

atio

n Si

gnal

In

tens

ity (

mV

)

1553 1554 1555 1556 1557

-80

-60

-40

-20

Wavelength (nm)

Inte

nsity

(dB

m

-50 -25 0 25 500

10

Delay (ps)Aut

o-co

rrel

atio

n Si

Inte

nsity

( m

V

(2) Gloop≥ 0dBnear threshold

(1) Gloop< 0dB-30

-20

-10

0

put o

ptic

al p

ower

, Pou

t [dB

(2)

(3)

(1)

+0 7

7.2dB

0

1

2

3

4

5

6

Out

put p

ulse

wid

th, τ

[ps]

easu

red

wid

th o

f pul

se, τ

(p

0

1

2

3

4

5

6

Out

put p

ulse

wid

th, τ

[ps]

easu

red

wid

th o

f pul

se, τ

(p

solid curve: 10 GHzdashed curve: 40 GHz

13Ultrafast Optical Logic Lab., UECUltrafast Optical Logic Lab., UECUltrafast Optical Logic Lab., UECUltrafast Optical Logic Lab., UEC

Delay (ps)

-10 0 +10 +20Pulse loop gain, G loop [dB]

Out

p +0.7 0 1 2 3 4 5 60

MZI's delay time, Δt [ps]Delay time in DISC, Δt (ps)

Me 0 1 2 3 4 5 60

MZI's delay time, Δt [ps]Delay time in DISC, Δt (ps)

Me d s ed cu ve: 0 G

Flip-Flop (Eindhoven, Tsukuba)XOR, AND, 2R/3R, etc.

fundamental research of “gates”

single-longitudinal-mode mode-locking, 2008

Clock oscillatorClock oscillator (UEC)

precise only-one-mode lasing(with using high-Q etalon filter designed by JAE Japan)

precise, only-one-mode lasingout of Δ10-MHz-spacing modes.

(Nakamoto, et al. OSA-NANO 2008)

Original features of this scheme of ours: 500 GH BW b l ti i t ti ibilit ( tl ) 500-GHz-BW comb, low power consumption, integration possibility (presently) precise optical frequency, fopt (locked to external DFB source, fopt) precise repetition frequency, fR (locked to dielectric etalon’s FSR, fR)

i b l h f (l k d di l i d l i Δ )

1419Ultrafast Optical Logic Lab., UEC

precise comb envelope shape, fBW (locked to dielectric MZI delay time, Δt)

status of Modeling-research (optical gates)

[2] Latest optical gates and memoriesfairly-good

Ueno et al., 2001-2002ECOC, IE3 PTL, IEICE Trans.

Subject: about the useful correlation between

from the hybrid-integrated SMZ device

Subject: about the useful correlation betweensensitive dependences of waveform and spectrum, on the optical phase bias, ΔΦB

meas.meas. calc.

C l i

vefo

rms

log

scal

e Conclusion:

we’ll be able to feedback-control

wav

in th

e the 160-Gb/s Demuxby monitoring the “supervisory” spectrum

Ultrafast Optical Logic Lab., UEC

[2] Latest optical gates and memoriesfairly-good

status of Modeling-research (optical gates)

J. Sakaguchi et al., JJAP 2005 and 2008

subject: about one generic issue for DATA conv.in the original DISC design

calculated measured

s ale

avef

orm

se

log

sca

wa

in th

To solve this, within DISC scheme, we need a kind of imbalance factor between the two interferometer arms.

Ultrafast Optical Logic Lab., UEC

[3] Physics and potentials, of all-optical gates <speed, energy, size>

[3] Physics of all-optical gates

Quote: “All-optical semiconductor gate seems too complicated”--- Prof. Guifang LI, CREOL/UCF, USA.

Injection Current (mA)

Inversion-population semiconductors (SOA’s)

• nonlinear dependences → unsaturated gain G0, gain-saturation energy Psat, optical 3R/2R.

p p ( )

• linear dependences (many ways) → in gate speed, energy, size.

0.50

1.00

ift, Δ

Φ/π

42 GHz

84 GHz

168 GHz

Ueno et al., JOSAB 2002

“not so complicated !!”10 50 100 500

0.10

Injection current (mA)

Phas

e sh

i



j ( )

linear dependence

Physics ： Refractive-index modulations, etc.,due to excited-electron-hole-density modulations

K Tajima (NEC 1993)

[3] Physics of all-optical gates

Holes Excited electrons in conduction bandOptical output Z

K. Tajima (NEC 1993)J. Sokoloff (Princeton 1993)

Optical pulses Amplified pulsesOptical output X’

SOA

Excited holes in valence bandOptical input X

Optical input Y Step wise processes in generating optical outputs Z and X’t

Optical input Yand

optical acceleration Y or A

Step-wise processes in generating optical outputs, Z and X’.

(1) Stimulated amplification of optical inputs X and Y.⇒ (2) electron density is modulated (in materials).

⇒

Electrons

⇒ (3) refractive index is modulated (in materials).

material’s rise time= 100 fs (fall time is slower).

⇒ (4) optical phase is modulated (in optical signal X or Y).

⇒ (5) d l t d li ht h tl bi d b f t ’ t t

D i d l t i j ti (d bi t)

⇒ (5) modulated lights are coherently combined, before gate’s output.

⇒ (6) new optical outputs, Z and X’ are generated.


Drive energy= dc-electron-injection energy (dc-bias current)

Speed of gates[3] All-optical gates/ speed

faster than material-relax. speed,after optically accelerating the material

Holes

Origin: R. Manning (BT), EL 1994

Optical output Z

Optical pulses Amplified pulsesOptical output X’

SOA

t

Optical input XOptical input Y

andoptical acceleration Y or Aoptical acceleration Y or A

(seed cw or clk)

Drive energy= dc-electron energyElectrons

• For gates, this acceleration is equivalent to “faster materials”

gy gy

For gates, this acceleration is equivalent to faster materials• Energy consumption is equivalent to them, as well.

experimentally recognized.(Sakaguchi, et al., Opt. Express 2007)


19

Speed of gates, more simply Sakaguchi, et al., Opt. Express 2007Ueno et al JOSAB 2002

[3] All-optical gates/ speed

material level: nonlinear dependence

Linear dependences after optical acceleration

Ueno et al., JOSAB 2002

Principle of acceleration (in SOA)

50

100

1/τ e

ff (G

Hz)

B#3, Iop = 200 mA

measured calculated|excited>

(quasi fermionic)

material level: nonlinear dependencePrinciple of acceleration (in SOA)

-50 -40 -30 -20 -10 0 +10 +20 +30

5

10

Rec

over

y ra

te, (quasi-fermionic)

relax. due to stimulated-emission(with seed cw or clock)

= optical accelerationHolding-beam input power (dBm)

P OP

|

continuous pump(efficient)

= optical acceleration

gate performance: linear dependences show-up.energy per bit ∝ speed (due to Joule loss)

helpful for designs and experiments

103

104

onsu

mpt

ion,

Φ

NL=

0.3

π)

B#4 1100 μm

B#3 500 μm

measured|g>

(transparent-state)

102

103

ic d

c po

wer

c(m

W, @

Δ Φ

optical phase mod. ∝ electron dens. mod.ti l h d ∝ i t ti l th

several other linearities::

20

10 100 1000

Elec

triBandwidth, B (GHz)


optical phase mod. ∝ interaction lengthoptical phase mod. ∝ electron-photon overlap

Speed of gates numbers at operating point (200Gb/s)[3] All-optical gates/ speed

Principle of gate = electron-pump

w/ nonlinear pol. rot.S. Nakamura et al., Appl. Phys. Lett. 2001J. Sakaguchi et al., Opt. Comm. May 2009

Optical pulses Amplified pulses

Holes

168G wavelength conversion (2000)SOA

toptical input (30fJ/bit)

0.4

0.5

0.6

a.u.

)

0 4

0.6

a.u.

)

168G input data 168G output data

Electrons

acceleration (cw, 100 fJ/bit)

-50 0 +500.0

0.1

0.2

0.3

D l ( )

Sign

al (a

-50 0 +500.0

0.2

0.4

D l ( )

Sign

al (

Delay (ps) Delay (ps)

material > 60ps gate recovery < 6ps

l t ti 1×107 electrons/bit

dc-electron energy

・ electron consumption= 1×107 electrons/bit・ electric-energy consumption= 3 pJ/bitnearly-regardless of material’s speed (in this regime)


[3] all-optical speed energy

Energy-efficiency of gates new potentials in near-future (--2025)

Other institutes: packet-routing, buffer memories, integrated 2R-subsystem.300 pJ/bit30 pJ/bit

(previously)

・ blue-shift filter Nielsen et al Opt Express 2006

Our group (UEC): following directions (for energy-saving)

(previously)

・ nonlinear-polarization rotation Sakaguchi et al., Opt. Comm. 2009

blue shift filter Nielsen et al., Opt. Express 2006

・ spectral synthesis scheme going on (Nishida et al IEEE LEOS 2009)

3 pJ/bit(present)

・ spectral-synthesis scheme going-on (Nishida et al., IEEE-LEOS 2009)

・ (non-deg. ->) degenerate-scheme going-on[free from polarization-insensitivity requirements]

（UEC-NICT collaboration, FY2007-）

・ QW-band-engineering, for enhancing refractive-index mod.

[ ee o po a at o se s t ty equ e e ts]

not yet started but, from near-future0.3 pJ/bit(near future)

• efficient pump (optical) 100-fJ to 30-fJ

(e.g., ACQW)further


• low-dim., surface plasmon, nano-photo (far future)

[3] All-optical gates/ size

Size of gates interaction L, new potentials in near-future (--2025)

energ s ppl of 3 pJ/bit that is 1×107/bit of electrons in interaction length 1 mm

dc-electron pump (at present)

Volume density of excited electrons > 2×1018cm-3 ?energy supply of 3 pJ/bit, that is, 1×107/bit of electrons, in interaction length= 1 mm

exeprimentally: stock= relatively dilute (2×1017cm-3) ！Sakaguchi et al., Optics Express 2007 through present+600

+800

+1000

E-E c

(meV

)

conduction band

0 2.0 4.0 6.0 8.0 10.00

+200

+400

Density of electron states, Ne(E) x10+17(cm-3.meV-1)

Ener

gy, E dilute electrons

(non-degenerate)

n= p= 2×1017cm-3

flow (modulation)= less efficient (at present)

+800

+1000

eV) 10-times more electron density （2×1018cm-3 ）,

conduction band

one of new potentials

Characterstic temperatures0 2.0 4.0 6.0 8.0 10.00

+200

+400

+600

+800

Ener

gy, E

-Ec (

me

dense electrons(degenerate)

y （）, 10-times shorter gate length (L= 100 μm)

(Hetero-barrier energy seems not enough, at present)


(AlGaInP/GaInP laser, 1993)Density of electron states, Ne(E) x10+17(cm-3.meV-1)

volume density of excited electrons (cm-3)[3] All-optical gates/ size

Metals1E22

AlAu, Ag, Cu

Plasmonics(very lossy) volume density

1E20Ti:Al2O3, Nd: YAG

Mn: GaAsSpintronics“half-metal”

Semi-metals

1E18

2 3

super-dense electrons in III-V, 1018～1019

1E16

Er:SiO2III-V in all-optical gates (presently)unexpectedly dilute, 2×1017

Semi-conductors(un-doped, intrinsic)

1E16

1E14background levels

partially after Charles Kittel,Introduction to Solid State Phys.,

1E14


y1953-1996

24

Super-high-density electron-confinement,with new hetero-barrier systems

[3] All-optical gates/ size

Higher-Hetero (proposal)at present

future

presently Direct-Modpump laser

0 98

0.87 μmGaAs/AlGaInP

(920 meV)

p

50-GHz direct-mod.

g(e

V)

2.0

2.5

1.55 μm 1.31 μm

1.31 μmAlGaInAs/InP

(750 meV)

(DR laser) 0.98 μmInGaAs/AlGaAs

(760 meV)all-optical’s

(920 meV)

920 meVhetero-barrier

nerg

y, E

g

1.5

μInGaAs/InP

(550 meV)

μInGaAsP/InP

(400 meV)

(750 meV) 920 meV(= kBT×35)

dgap

En

1.0 550 meV

*) 伝導帯の状態密度も

Ban

d

0 0

0.5) 伝導帯の状態密度も

やや増加する。

そして、電流注入⇔光注入を、分離。


0.0

[4] Optical micro-processor, near Year 2025[4] Optical MPU near 2025

near-future <speed=300G, energy=0.3pJ/bit, size(interaction)= 100μm>

Electronic Optical processorElectronic

Specification intel 4004 Present Near Future

Demo Year Year 1971 Year 2000 2010 Year 2025

Optical processor

Demo Year Year 1971 Year 2000-2010 Year 2025

Speed 500 kb/s 200-300 Gb/s 300 Gb/sE ( bi ) 3 10 J /bi / 0 3 J /bi /Energy (per bit) 3-10 pJ /bit/gate 0.3 pJ /bit/gate

500×500 μm2

(w/ 100-μm interaction)1,000×3,000 μm270×70 μm2Size (per gate)

2,300 gates(6 chips on 3-inch wafer)

several2,300 transistersNumber of gates (per chip)

Energy dissipation (per chip) 200 Watt


Optical processor 4004

Relative performance of optical-processor 4004

[4] Optical MPU near 2025

Zetta1 E+20

1.E+21 Tr・Clk Product

Exa1.E+17

1.E+18

1.E+19

1.E+20

Peta



NEC-Sun Tsubame ->

1 E+13

1.E+14

1.E+15

1.E+16

intel 4004(1971) Core 2 Quad

Tera

1.E+10

1.E+11

1.E+12

1.E+13

Pentium Pro(1996)

Pen 4(2002)

(1971) Core 2 Quad(2007)

GigaCray-1 ->

1.E+06

1.E+07

1.E+08

1.E+09

80386(1985)

(1996)

8086(1978)

Earlier statements:• already relying on parallel-processing structures,

which are probably pushing-uptheir electric energy consumptions

1.E+04

1.E+05

1.E 06

1960 1970 1980 1990 2000 2010 2020no. of Trs. times clk freq. 10x every 4 yrs.


their electric-energy consumptions.• Increasing demands in 2010-2050 are

FLOPS-type demands or MIPS-type demands??


FLOPS 10x every 4 yrs.Instructions per sec. 10x every 7 yrs.

出典： FLOPSは市販PC計測結果、MIPSはwikipedia資料。

Relative performance of optical-processor 4004


Zetta1 E+20

1.E+21 Tr・Clk Product

Exa1.E+17

1.E+18

1.E+19

1.E+20

Optical processor (estim.)alternative to Moore’s law number-integration

Peta



NEC-Sun Tsubame ->

1 E+13

1.E+14

1.E+15

1.E+16

Core 2 Quadintel 4004

(1971)

Tr・Clk matches to Pentium Pro(1996)

Tera

1.E+10

1.E+11

1.E+12

1.E+13

Pentium Pro(1996)

Pen 4(2002)

Core 2 Quad(2007)

(1971)

(M)IPS matches to Core 2 quad (2007).

GigaCray-1 ->

1.E+06

1.E+07

1.E+08

1.E+09

80386(1985)

(1996)

8086(1978)

FLOPS will be weak.

Power consumption: 200 Watts

1.E+04

1.E+05

1.E 06

1960 1970 1980 1990 2000 2010 2020no. of Trs. times clk freq. 10x every 4 yrs.



FLOPS 10x every 4 yrs.Instructions per sec. 10x every 7 yrs.

出典： FLOPSは市販PC計測結果、MIPSはwikipedia資料。

sample-spec. numbers of optical-80386250×250μm2 o-Tr’s


e-80386 e-Core2_quad optical 80386(32bit) (32bit/64bit) (32bit)

250×250μm o-Tr son 6” GaAs wafer

(32bit) (32bit/64bit) (32bit)

1985 2007 2025unit

1 Clk Hz 1.20E+07 2.60E+09 3.00E+112 Tr 275 000 2 000 000 000 275 0002 Tr 275,000 2,000,000,000 275,000

Clk*Tr 3.30E+12 5.20E+18 8.25E+16Clk*Tr (relative) 6.35E-07 1.00E+00 1.59E-02

3 Flops 1 2E+05 4 0E+10 -3 Flops 1.2E 05 4.0E 104 (M)IPS, measured and estimated 1.10E+07 6.00E+10 2.75E+11

(M)IPS (relative) 1.83E-04 1.00E+00 4.58E+00

5 power consumption Watt 3 80 2.00E+0420 kW

p p

6 energy / instruction J 2.73E-07 1.33E-09 7.27E-08energy / instruction fJ 2.73E+08 1.33E+06 7.27E+07energy / instruction (relative) 1 5.45E+01

7 energy / clk J 2.50E-07 3.08E-08 6.67E-08energy / clk fJ 2.50E+08 3.08E+07 6.67E+07energy / clk (relative) 1 2.17E+00


optical-processor 80386 (with 300,000 gates)advantage of serial-processor


i l 80386 (

serial-processor

advantage of serial processorlatency, synchro issues 1/100 or less

programming costs, risks 1/100 or less

1.E+12

1.E+13

Core2-quad(2007)

• Clk: 100x

optical-80386 (estim.) technical efficiency 100x / 15 yrsparallel

1.E+10

1.E+11

e-80386(1985)

(2007)

• MIPS: increases

• energy per clock: comparable (70 nJ/clk)( ll l ti t C 2 d)

1.E+08

1.E+09 (all relative to Core2-quad)

c

1.E+06

1.E+07

Possible to integrate??1.E+05

1980 2000 2020 20401980 2000 2020 2040Possible to integrate??

250×250μm2 o-Tr’s

on 6” GaAs wafer

⇒ 275,000 Tr’s (=80386)

Previously,

Tr number×Clk: 100x / 8 yrs

FLOPS 100x / 8 yrs

1960

Q t l t i hi t

Opto-electronics history


MIPS 100x / 15 yrsQuantum-electronics history

Summary (ssdm 2010)

If anybody likes to keep this talk’s slides, please email me!

• Impacts of ICT-related energy consumptionse.g.: energy supply to Data centers in USA: 10 nuclear reactorsg gy pp y

heat energy from one server rack: 20-kW level.many-folded parallel-data-processes will the best for all applications, thru. 2050 ?

• <speed, energy, size> of all-optical gates, at present: <200G, 3 pJ/bit, length, 1 mm>

• <speed, energy, size>, 2nd or 3rd generation: <300G, 0.3 pJ, 250 μm2>

• in Materials Research (semi-classical quantum):optical acceleration (incl. gate scheme),electron-photon interaction (little studied),higher-density excitations (w/ larger hetero-barrier).

ti l 4004 MIPS bl t C 2 d El t i 200W• optical-4004: MIPS, comparable to Core2 quad. Electric energy, 200W.• optical-80386: 300,000 gates. Energy per clk, comparable to Core2 quad.

(this will probably save energy and costs, for a group of serial-process-oriented tasks.)


an alternative to 40-year-long Moore’s law

Thanks to Co-authors and Collaborations

Co-authors:Co authors:Jun Sakaguchi, PhD course, UECRei Suzuki, master course, UECTakashi Ohira, master course, UECTakashi Ohira, master course, UECRyoichi Nakamoto, master course, UECFellran Salleras, postdoc, UEC (from ETH, Laussanne)Mads L. Nielsen, visiting researcher from Technical Univ. Denmark (DTU), g ( )Kohsuke Nishimura, KDDI Labs.Naoya Wada and Satoshi Shinada, NICT Labs.and several more.

Collaboration:Jesper Mork, Technical Univ. Denmark (DTU)Juerg Leuthold, Karlsruhe Univ.Kiyoshi Asakawa, NIMS (Univ. Tsukuba)Shigeru Nakamura, NEC Labs.Harm Dorren, COBRA, Eindhoven Univ. TechnologyJian Wu, Beijing Univ. Posts and Telecoms (BUPT)K.V. Reddy, Pritel Inc., Chicago USA


Alistair Poustie, CIP Technologies, UK, and several more institutes.32

Documents

optical semiconductor devices to Year 2025to Year …If anybody likes to keep this talk’s slides, please email me!Roadmap of ultrafast energy-saving optical semiconductor devicesto