Ambient Intelligence: A Gigascale Dream Facing Nanoscale ... · Ambient Intelligence: A Gigascale...
Transcript of Ambient Intelligence: A Gigascale Dream Facing Nanoscale ... · Ambient Intelligence: A Gigascale...
Sensorsysteme Dresden 1
Ambient Intelligence: A Gigascale Dream Facing Nanoscale Realities
Hugo De ManProf. K.U.Leuven
Senior Research Fellow [email protected]
© imec 2005
?
Giga-Scale Complexity Nano-Scale Realities
AmI Systems Atoms
10nm
Sensorsysteme Dresden 2
The Ambient Intelligence Dream
Secure, trustworthy computing and communicationembedded in every-thing and every-one.
A pervasive, context aware ambient, sensitive and responsive to the presence of people
Sensorsysteme Dresden 3
Device Classes for AmI
Stationary Nomadic Transducer
Am
bient Body
Home
OfficeCar
“UPA”
TCRF
TCRF
TCRF
UMTSWLANWPANWBANDA/VB
PDADSCDVCMP3GPSHC…
Hear, See, Feel, Show…
“More Moore” “More-than-Moore”
MIMO
Internet IPv6
100Gop/s 10 Gop/s
1Watt 100mW
‘Milliwatt’ (battery, f.c.)
10 Watt
‘Watt’ (mains)
1Top/s 10Mop/s
Gb/s
0.1Gb/s
kb/s
100µW
energy
energy
energy
(ambient)‘Microwatt’
Sensorsysteme Dresden 4
Outline
More Moore: ultra complexity
Bridging the gaps between systems, software and atoms
More than Moore: ultra creativity
The interface challenge
Sensorsysteme Dresden 5
Design Metrics for Ambient Intelligent Digital Devices (isscc’05, 1.2)
• Power Efficiency (PE): 10 to 200 GOPS/Watt
• Part cost < 10€ NRE cost > 50M€PE two-orders-of-magnitude > GP microprocessor
at 1/20 th of part cost
• Flexible by (downloadable) embedded softwareFrom Hw centric ASIC to Sw centric PLATFORM
• Real-time stream based processing (MM, SDR)Memory intensive (> 70% area, power)
Sensorsysteme Dresden 6
Two Gaps between AmI-Dreams and Nano-Scale Realities…Software centric AmI specs
7th heaven of software
Hell of nano-scale physics
500M tr Platform architecture
500 processors50MB memory
1 Watt
xx nm Si IP Blocks
PlatformDesign
IPCreation
CMOSScaling
SystemDesign
SDR
MiddlewareAPI’sRTOS
Architectural gap
*Techn. Aware Design”
TAD*
ESL
UncertaintyMetrics!
Physical gapDFM
Sensorsysteme Dresden 7
1. Hell of Nano-Scale Physics
• Leakage power starts to dominatelarger gate delay than scaling for performance predicts
• Voltage headroom shrinksmakes A&RF guys deeply worried (sub 1 Volt circuits)
• Interconnect claims the first role…challenging timing, power, synchronism, signal integrity
• Increasing device variabilityjeopardizes predictability and yield, affects design process
• New device architectures needed<65nmImpact layout style, need for DFM and IP library design
Sensorsysteme Dresden 8
Leakage Power IssuesPsdleak ~W.Tj
2exp(-qVTH/nkTj)
PsdleakIon~ (VDD-VTH)1.5/tox
Hence “LateCMOS” devices below 65nm?
Pgateleak
tox
Sensorsysteme Dresden 9
“LateCMOS” Device Architectures below 65nm?
• Gate leakage: Hi-k, metal gates (when?)• Mobility keeper: Strained Si (now)• SD leakage: Hi VT ⇒Hi-VDD ⇔ dyn power
Lo-VDD ⇔ variability, ↑td , ↑ //Better SCE mgmt->UTBFDSOI, Finfets
• Ion boost,SCE ↓ Multi-Gate, FINFET, DGFDSOI?
2005 2015
L=35nm
SiGe
L=35nmL=35nm
SiGe
strainstrainHfSiON2
high high --kk
NiSi
FUSIFUSI
metal gatemetal gate
CNT??UTBSOI 32nm?
BUT Late CMOS diversity must percolate to system level…
Sensorsysteme Dresden 10
Ultra-Low Vdd for Sensor Intelligence
Leakage Ener gy ( over 1ns) Vs. Vdd
0
5
10
15
20
25
30
0 0. 2 0. 4 0. 6 0. 8 1Vdd ( V)
Elea
kage
(aJ
)
* Data for 4-bit adder (130 nm CMOS) obtained using Monte-Carlo simulations over process distribution, Vth=0.24V
Active and standby power consumptions reduce quadratically with Vdd scaling.
7X
Swi t chi ng Ener gy Vs. Vdd
0
60
120
180
240
0 0. 2 0. 4 0. 6 0. 8 1Vdd ( V)
Eswi
tchi
ng (
fJ)
9X
Source: Qin, Rabaey UCB
Switching Energy vs. Vdd Leakage Energy (1ns) vs. Vdd
Sensorsysteme Dresden 11
Price to Pay at Ultra-Low Vdd
Del ay Var i at i on/ Mean Vs. Vdd
01020304050607080
0 0. 2 0. 4 0. 6 0. 8 1Vdd ( V)
tp V
aria
tion
/Mea
n (%
)
Delay Mean Value
• Delay variance at low voltages very dramatic
Delay Variance/Mean
* Data for 4-bit adder (0.13 µm CMOS) obtained using Monte-Carlo simulations over process distribution, Vth=0.24V
Delay Mean Vs. Vdd
00.5
11.5
22.5
33.5
0 0.2 0.4 0.6 0.8 1Vdd (V)
tp(u
s)
• Mean performance degrades exponentially below threshold
Source: Qin, Rabaey UCB
Concern over data storage: Memory Functionality / Stability
Sensorsysteme Dresden 12
Different CMOS for Different People…
For Microprocessor, Servers
@100 Watt active power, 20W leakage acceptable ⇒ low VTH , high Ion, ABB to manage leakage yield
BUT for AmI…
Leakage mWatt nodes <10mW, µWatt sensor nodes<10µW⇒ Scaling for low leakage at expense of gate delay⇒ More GOPS from more transistors faster ones⇒ Wires more important anyway…change layout style
Sensorsysteme Dresden 13
Scaling for Low Leakage AmI
( Courtesy: Philips Techn. Modeling Group)230 kgate Standard Cell Layout for 10pA/µm sd Leakage
BULK CMOS
SG
DG
FDSOIMUGFET
0
0.2
0.4
0.6
0.8
1
180 130 90 65 45 45 nm
WIRESCELLS
Avg
. Ene
rgy
/ Cyc
le (R
el.)
Techn. Node
1.8V
1.2 1.2 1.2
0.75
1.2
Lo-k dielectric barrier materials?Keep wires further apart Localize interconnectBack to regular layout (EDA!)
CuCu Cu7.04.0
Low k
CuCu CuCu
Low k
Hi-k
2.0
BEOL technology or layout tool work neededBelow 65nm new device architectures needed
Sensorsysteme Dresden 14
Global Interconnect Issues
• Global Interconnect delay does not scale with logic• Synchronous zones 45nm @ 500MHz < 2*2 mm2
• Interwire capacitance jeopardizes timing closureSS SSVV VVGG
VDD
SignalVss
Source:Khatri DAC’99
• CMP Copper Thickness Variation up to 40%
Source: Preagasus Delay Variation
Sensorsysteme Dresden 15
Random Dopant Variability …
35nm Bulk CMOSTransistorMin. size
Asenov, ESSCIRC’04
100
LeakageVariability!
σVT = Ao(L)/(WL)1/2
Sensorsysteme Dresden 16
Variability Impacts Parametric Yield
Low Standby Power CMOS; Min size transistors
0
0.2
0.4
0.6
0.8
1
1.2
1.4
180 130 90 65 45 32Techn. Node(nm)
VDD
VT
VT+3σ∆VT
VT-3σ∆VT
BULK
VDD, VT (V)
Headroom
noise
0.5
1
1.5
2
2.5
3
180 130 90 65 45 32
Gate Delay Variability
BULK
VT+3σ∆VT
VT-3σ∆VT
Slow
Leak
y
Techn. Node(nm)gate
N
VDD
DGFDSOI DGFDSOI
Yamaoka VLSI‘04 Wfin
Lg
Hfin
Sensorsysteme Dresden 17
Living with VT Variability in Logic
• Use min Lg only for critical delay path• Use large W when matching is critical• Use large logic depth (less pipe stages, slower
clock!)Ntdo
Fcl < 1/ Ntdopd
fpd
f
tdo
N=1
N>>1
Ntdo
Ntdo
tdσσ =
σVT = Ao(L)/(WL)1/2
Sensorsysteme Dresden 18
Long Lg Usage Pattern in µP
Sizing matters for :
-Leakage -Variability
Mitigation!
Source: Intel ISSCC 2006 p05.3
SRAM!
Computation
Sensorsysteme Dresden 19
Variability Impact
• A new dimension to design space (P-A-T-Y) • Statistical design, device sizing for yield crucial
• SRAM first victim…else VDD scaling problematic• Below 65nm new device architectures needed?
45nm FinFet SRAM IEDM’04
0.31µm2 6T SRAM FinFETs, FUSI gate
imec3D SRAM
Samsung isscc’05Asenov Esscirc ‘04
Sensorsysteme Dresden 20
Reaching the end of UV litho => DFMMask design cost ↑
OPCPSM
SRAF: Sub Resolution Assist FeaturesSB : Scattering Bars
Rule Based DRC runs out of steam: 2000 pages for 22nmTowards Model Based hot-spot detection and correction Booming DFM EDA industry Towards Restricted Design Rules <45nm (Regular layout)32nm => EUV?
Sensorsysteme Dresden 21
Sys+
Tech
Software dominated
2005-2020 Not Business as Usual!
1980 1990 2000 2010 2020
Cum
ulat
ive
Inte
rdep
ende
ntC
halle
nges
500 250 180 130 90 65 45 32 22nm … ???
ManualDesign
T+S Happy Scalingtech sys
RTL HwIP PlatformAREA TIMING
Variability
YIELD
5 3.31.8
1.2
Gate LeakageHi-k, MG
LateCMOS FINFETFDSOI
M-MetalgateHi-k,Lo-k
StrainSiGeCNW&T
RFNvmem
SUM…
NewDesign
Dynamic PowerLow Cost, Power, TTM > Fcl
1 0.5 0.2V
POWER
SD LeakageULV
Sensorsysteme Dresden 22
2. The Architectural Gap
“Managing Giga-Complexity ”
Create SL methods, tools and skills for designing flexible, yet power-efficient platforms
(1 B uncertain devices + >10M lines of code)and for mapping real-time MULTIPLE SW applications on
them at lowest NRE cost
Challenges• Overcome Power-Flexibility Conflict• Predictable (SW) application mapping• While at the same time dealing with
the CMOS process ®evolution
Sensorsysteme Dresden 23
Power-Flexibility Conflict
32 bit IntrinsicPE
2 1 0.5 0.25 0.13 0.07feature size(µm)
1000
100
10
1
0.1
0.01
0.001
Power Efficiency (GOPS/Watt)
AmI
SwClean
//
Reconfigurable // computingMuxed data paths
IS Computing
[Claasen isscc’99]
FGCG
µP
VDD<1VVIP
SiHiveCoolflux
AdresFeenecs
Sensorsysteme Dresden 24
How to Reconcile PE and Flexibility?
• Cleaning software for low powerMinimize # memory transfers, operations = energyRegularize, vectorize code, min. data-sizeFactor 5 to 10 lower power but tools and skills needed
• Exploit parallelism to reduce powerBetter to use 20 processors @ 100MHz than 1@ 2GhzBut urgent need for design methods, tools and skills to support Hw/Sw co-design (devil is in the software!)
• Exploit task level dynamism for sub 1 Volt opJust enough VDD, Fcl, IL
Sensorsysteme Dresden 25
The Devil is in the Software and the NRE cost…
Software productivity increase 5.8%/y <-> 58%/y Moore
Hw dependentReal TimeReliability
DependabilityConcurrency
HeterogeneousParallel
HardwarePlatform
Merging of hw and software cultures is people management …changing people is harder than installing and operating a new litho machine!
2 y/ tech node7y / methodology
Sensorsysteme Dresden 26
But System Design must change…
Better than worst case design else loose advantage of scaling
Design reliable systems from uncertain components
Allow errors to happen!Shannon
1948
Sensorsysteme Dresden 27
Ways to Cope…
1. Adapt VDD,fcl,IL at run time to whatever mother nature provides and to dynamic nature of software tasks
2. Use tile based GALS architectures, networks-on-chip, and error correcting on-chip communication techniques
3. Back to regularity (Litho, context)
RapidChip
Sensorsysteme Dresden 28
Run Time Adaptive Design
Processor1
Processor2
Processor3
t
t
t
2
2
2
1
3
1
1
3
3
mon
itorin
g
cycle budget CB
Min Σ Eit1+t2+t3 < CB
SoftwareRTOS
HardwarePlatform
Run-Time Controler
TasksCB
Optimal SchedulingFcl ,VDD,VBB
Courtesy: Catthoor isscc03 9.1
IPMonitorstd, IL, T
Dynamic Task scheduler/assigner
Another new design dimension to mitigate leakage and variability and improve yield and performance
Pareto set of optimal task mappings
Sensorsysteme Dresden 29
<1GHz<4mm2
<1GHz
45nm
GALS Platform of Future + NOC
I/O
peripherals
3-D stacked m
ain mem
ory
RouterBusBasedMultiPr
ME
LocalMemory Hierarchy
ErrCD
NIP
Power
TestM
gmt
SYNCH TILE
Mapping V,VT,fcl
IL,T,td
Middleware, RTOS, APIPA- Run-Time ControlerApplications SW
Clean
30Mtr
ASYNCH Low swing Network-on-chip
SS SSVV VVGG
Sensorsysteme Dresden 30
More Moore = more than more of the same…
• Greater challenges than ever before: Managing giga-complexity faced with CMOS diversificationNew design methods, tools and skills urgently needed
• To reconcile the hell of physics with heaven of AmI dreams
But easier to install equipment than to change people’s mindNew processing nodes 2 years, new design methods take 7 years,
• Need for sharing multi-disciplinary R&D costFrom material science to computer science and societal
engineering…Open innovation alliances from software to process development
• Not for the faint of heart: few platforms by industry-R&D alliances and disciplined armies of engineers…
Sensorsysteme Dresden 31
But there is “More than Moore”…
• Key to AmI: interfacing to the ambient and to the consumer…
• Needed are the simplest, cheapest intelligent micro-transducer nodes…
• Combining a plethora of technologies…MEMS, Displays, Energy scavenging, SiP, biosensors,
security, ULP radio protocols, ULP computing…
Sensorsysteme Dresden 32
More than Moore Challenges
The art of ingenuity
• Get to the ultimate limits of Miniaturization (<1cm3)Cost (< 1€)Power ( < 100µW)
• Design for utmost simplicity • Interact with non-E world• A micro-system node in ad-hoc network
Sensorsysteme Dresden 33
Heterogeneous Integration
The key art of More than Moore design:need for the renaissance engineer
RF-ID tag
Wheel speed Phone camera 3D accelerometer
Inkjet head Ultra-filter
RF-MEMS switch
Lab-on-a-Chip Nano-syringe
GSM frontend
fluidicspassives
RF
high-voltagesensors &actuators
fluidicspassives
RF
high-voltagesensors &actuators
Sensorsysteme Dresden 34
Sensor Node Challenges
<10kb/s1%
DSP&storageSecurity
MAC
RFNon-EWorld
Sensor CE-ADC Processor PicoRadio
20µW 20µW 40µW 20µWAvg.
Power80 Mops 2nJ/b
Energy sensorPower Mgr
< 500mV CMOSULP radio
Above IC MEMS, passivesGrain size packaging
100 µWAvg power
Ambient
Sensorsysteme Dresden 35
Sensor Radio’s
ProtocolµC
2.4 GHz Radio
ApplicationProcessor
(EEC)
Solar Cell
Antenna
(a) IMEC 1.4cm3, SiP Mote for EEC, ECG , 500µW@1%, 400b/sec
3D stack
(b) UCB PicoBeacon Tx [21.4]1.9GHz, 400µW, 5kb/s
2.4*3.9cm2Battery/
powermgmt
1 2 4 GHz
500MHz-70
-50
-30 dB
(c) IMEC 0.18µ 0.25 mm2 UWB Tx0.5nJ/bit @ 10 kb/s
BatteryPowerMgmt
Sensorsysteme Dresden 36
More Moore and Analog/RF
• Scaling: squeezing the analog/RF designerMatching, voltage headroom, substrate noise
Inductors / passives don’t scale (Above IC WFP, SIP)Go for dirty RF/Analog, correct errors by digital
10pF == area of 8000 gates @ 45nm10 bit ADC == energy of 4 M gates @ 45nm
• < 90nm useful for > 10 GHz trend
CourtesyRuthenbar CMU
Sensorsysteme Dresden 37
Above IC MEMS
Poly-SiGe
CMOS circuits
MEMS< 450oC
CMOS
MEMS
Yaw Sensor + ElectronicsCourtesy: BOSCH-PHILIPS-IMEC
Sensorsysteme Dresden 38
Above IC Integration Examples (IMEC)
Above IC InductorsHighest Q, high density
Above-IC CapacitorsUp to 10 nF/mm2
Above-IC digital Interconnect TRL lines
>10 Gbit/channel
5.5GHz ESD protected LNA on 90 nm rf-CMOS
Lowest noise
5GHz VCO on 90 nm rf-CMOSLowest power & phase noise
24GHz Single-stage LNA on 90 nm rf-CMOS
Highest frequencies
Sensorsysteme Dresden 39
Convergence on the Nano-Scale…
cmµmnm
BIOTECH
transistor
x109
NANOELECTRONICS
31nm31nm
Fromherz
Sensorsysteme Dresden 40
Interfacing Bio to Nano-Electronics
electrical
action potential
chemical
neurotransmitter
© IMEC 2005© IMEC-HUJI 2003
© Infineon 2003
© P. Fromherz 2005
© P. Fromherz 2001
ULP Interpretation-control-transmission
Sensorsysteme Dresden 41
4 NSF-SRC Questions for the Future…
HOW CAN WE COPE WITH INCREASED:
1. Complexity (NRE cost)2. Uncertainty (variability, reliability, noise,
workload…)3. Diversity (heterogeneous MTM systems)4. Adaptability (materials, device
architecture, 3D…)
ALL AT THE SAME TIME (Παντα Ρει)
Sensorsysteme Dresden 42
SO:
We need to lock nano scale device, circuit and hardware /software architects together under the same roof and
address strategic application domains
And take time to talk to each other!!
Collaborative multidisciplinary research needed between Academia-industry-research institutes
Sensorsysteme Dresden 43
THANK YOU