Confidential
Transcription
Confidential
2007 “Design Activities at LETI : New developments for Power Efficient SoC” Edith BEIGNE © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 1 Power Trends 2007 © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 2 Transistor Integration & Power dissipation 2007 © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 3 Power sources 2007 © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 4 Power components in digital CMOS Power due to switching capacitance Pdyn Power due to short circuit currents PSC Power due to leakage currents Pleakage 2007 P = α fCV 2 DD + I SCVDD + I leakageVDD © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 5 CMOS Power – With Subthreshold, Gate and Bulk leakage Dynamic Power + Active Leakage Leakage Power IBULK PMOS PMOS Network 1 NMOS 2007 0 IDYN Network 1 Network IGATE NMOS 0 ISC IGATE Network ISUB IBULK © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 6 Power Aware Strategies 2007 © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 7 Power aware strategies 2007 © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 8 Reducing Power : Real Application to a complex power aware SoC 2007 © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 9 Reducing Power : Real Application to a complex power aware SoC 2007 Chosen strategies © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 10 Multiple Threshold Voltages - Hardware Means Availability of Multi-VTH transistors in actual technologies Complete set of cells provided and characterized Logic levels remain unchanged Higher threshold voltage for leakage optimization 2007 Low threshold voltage to the most performance critical blocks Possible within cells © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 11 Power Gating - Hardware Means Per gate only modified cell library needed fine grained controllable high area overhead Per row 2007 only additional gating cell needed delay penalty hard to predict Per block no changes in module design needed temporal solution for IP reuse © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 12 Power Gating – Gate Boosting and Super Cut-Off Simplest : CCMOS single transistor, Vcut = VDD standby power determined by leakage floating through the gating transistor But : High on-resistance in active mode Not so small leakage in stand-by mode Gate Boosting In active mode, Vcut is increased (NMOS) or decreased (PMOS) 2007 Super Cut-Off CMOS In stand-by mode, Vcut can be increased above VDD (PMOS) or deacreased below VSS (NMOS) But : longer wakeup times Vcut limited due to possible oxide breakdown Mainly for subthreshold current Edith Beigné- June 22th, 2009 Confidential © CEA 2009. All rights reserved Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 13 Dynamic Voltage Scaling (DVS) - Architectural Means Supply Voltage and the effect on Circuit delay ≈ Dynamic 2007 asymptotical Delay Heavy increase in delay Heavy increase in power VDD (V DD − VTH )α quadratic Power 2 P = α fCVDD Leakage I subth Power W −β = αT ² e L exponential VTH T VTH = VTH 0 − f DIBL ( L ) ⋅VDD © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 14 Discrete DVFS - principle Use only some discrete available voltage values 2 or 3 values are necessary each one with its associated frequency Easy integration (low area) Hop between those discrete values VDD Hopping and Voltage dithering (c) 2007 © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 15 Reducing Power : Real Application to a complex power aware SoC 2007 ALPIN chip © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 16 Global architecture : ALPIN circuit An Asynchronous Low Power Innovative NoC implementing new design techniques to reduce dynamic and static power consumption on a 65 65 nm technology cu t_off IP U nit Vco re I N E ast IN We s t OUT E ast L S L S OUT We s t I N a Es t IN OUT We s t a Es t L S L S L S L S IP Uni t Vhig h Vnod e res et_ n L S L L se l e p S L Vc ore Node Ac ti v it y de t e c tio n IN No rt h OUT No rt h OUT Re s Local Clock Generator SL SL SL SL L L L L OU T a Es t OU T S o u h t N I Souh t L S L S L S L S OUT We ts IN Es at n V ode se l p No d e N I N ro h t O U T N o rt h A c it v y ti e dt ec t o i n L S L S L S L S L S IN We ts OUT Es at L S L S L S L S L S L S L S L S L S L S L S L S L S L S L S OUT S u o t h IN o Su h t L S OU T We s t I N a Es t N I OU T We s t a Es t OU T S o u h t L S L S L S L S N I We s t L S L S L S L S L S L S L S L S L S L S L S L S L S L S L S L S L S L S L S L S I N a Es t OU T R e s I NRe s Vhigh Vhigh OU T We s t Vh i gh S S S S IN e Rs I NRe s L S L S L S L S L S S SA L S L S L S L S SL SL S S L L S S S L SL LPM SL S L S SA L S L S L S L S L oc al Clock Generator e dt ec t o i n ck lo S L SL Vc ore A c it v y ti V core MEM SL O U T N o rt h Tl ow S L L S SL N ro h t Th g i h (xN) C L S L S L S L S S L L S L S L S L S L S L S L S L S L S re set_ n se l p SL S L n V ode No d e N I OU T R e s ck SL S L S Vh i gh L L S L S L S L S Vhigh L S L S L S L S - SL SL SL SL SL SL S L L Pow er Switches cu t_o ff S L L S SL S L S SA S L S L S L S L S Hopping Unit Clo S OUT E ast L S L S L S L S UCO cu t_o ff_cl k L PM S L SL Local Clock Generator IN We s t V low V high Vhigh L S L S L S L S L S L S L S L S Vhigh Vhig h r eset_ n signals SL Vc ore se l p Node Ac it v y ti e d t e tc o in I N E ast S L S L S L S L OUT S o u h t IN S o u h t 80c51 cut_ off sig nals dcdc_ctr l signal s L S L S L S L S S L L S L S L S L S L S L S L S OUT We s t OUT S o u th IN S o u th LPM rese t_n L S L S L S L S S L L S SL S L L S SL S L S L S L S L L S L S S L S L S L S L L S OUT We s t NoC perf ck Vh i gh n Vode IN No rh t OUT No tr h OUT Re s I N Re s L S L S L S L S L S S L S L S L S L S L S L S L S L S L L S L S L S L S OUT S o u h t IN S o u h t lo C S L S L S S L L S S L L SL S L Vhigh Vhigh Vh igh S SA L ocal Clock Generator e d t e tc io n I N Re s I N Re s S L S L S L S L s le p No d e Ac ti v ity OUT Re s L S L S L S L S rese t_n L S L S L S L S L S L S L S L S L S Vh i gh n Vode IN No r th OUT No tr h SL S L L ck SL S S L S L S L SL SL SL SL lo S L SL C S SA L PM Vcor e FHT2 L S L S L S L S S L SL S L SL L S SL k c SL S L L S L S L S L S L S L S L S L S L S SL S L se l p Node Ac it v y ti e d t e tc o in IN No rh t OUT No tr h OUT Re s lo SL rese t_n S L n Vode L PM S L SL N I C Vh i gh L DC-DC Conv. T arge t val ue Local Clock Generator Vco re S L S L SL SL SL SL S L S Co nt ro lled S wit ch Po wer f ilte r Cor rect or Vcor e S SA L S L S L S L S Automatic frequency scaling Use of clock generation re-programming to find the optimal V/F point of operation ck Local Clock Vcore G enerator A local fine grain Dynamic Frequency Scaling : 2007 Implementation of a local hardware controller to control transitions between Vhigh and Vlow Ensures smooth DVS transitions for IP safe computation SD mod ulator dcdc _c ontrol FHT1 NI S L o S L L S SL Cl rese t_n Pow er Switches NI LPM I P U ni t V core OFDM A local fine grain Dynamic Voltage Scaling : To l w Hopp ing Unit Hopp ing Unit T hi gh (x N) - V high Vlow UCO cu t_o ff_ cl k Power Switches L S L S L S L S L S L S L S N I Souh t L S L S L S L S To l w N I T hi gh (x N) N I cut_ off - Vh igh Vlo w Vhigh UCO cu t_o f _ cl k NI Each synchronous IP is an independant power and frequency domain ALPIN Architecture Thanks to pausable clock technique, IP unit continues its operation during DVFS phases Local fine grain power management can be executed during IP computation and communication independently from the others © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 17 Vhi gh UCO Th g i h( N x) tv i y c ti n o e p o N e d A c t iv t y d e t c ti n o N o rt h U O T N o r th NI S SL SL I R N e s R N e s N I W e s t E a s N I t L O U T E a s t S S L S L U e O T W s t L S E IN a s t L L S L S L S L S L S IN e W s t O T U E a s t S L S S L S S N T or N I oS hut ht de ct tvi y et ict on N o e d I O U T N rto h O U N T ro O U T o S tu h N I o S th A c t tv i y d c ti n o e t S L S S L S L O U T t L S L W e s t S S L N I L S S L W e s t t L S N I E No C S S S SL Vnode IN No rth O UT North IN East IN West OUT East LS LS LS LS LS LS LS LS LS LS LS LS O UT South IN So uth LS Level shifters + isolation cells O UT West LS LS LS Asynchronous Node © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential S L S L S L S S L S L S L Vhi gh sl eep Node Activity detection OUT Res LS Network Link S I Vhigh LS LS S L esl L N ort h O U N T or O TU oS hut N I oS hut ht E Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 18 S L S L S L S L Network Link L Vhigh LS LS LS I esR ep oN de A ct tvi y d et ict on L IN Res LS T es N Ias O TU E as S t L L S t L S L eW st L L S S L S L S L L LS LS LS V o n e d LS O U R N eW O U sTt N I LS us S V ig h h S LS LS LS LS ol oc ot LS LS LS ch yn As no ro pr S L SL L L LS S S S S S S L S L S L Vcore Vcore S SA LS ri ai C LS LS LS Local Clock Generator No LS dl_cfg us S SL SL SL SL S Vhi gh S L L Vhigh S L S L S L L t S L S L In Da p u t ta Hopping Unit LS pwc_hl Ou Da tpu ta t h nc Sy no ro ol oc ot pr LS k oc Cl LS IP Core Synchronous Vcore LS IP Unit Vcore LS Local Power Manager Vref Synchronous Vcore LS L L a s S LS D2 ?A Network Interface reset_n L t O U T E tu h + LS 2007 INIT : reset at Vhigh (1.2V) HIGH : Vhigh supply LOW : Vlow supply (0.8V) HOPPING : switch Vhigh / Vlow for DVFS IDLE : retention state at Vlow (no clock) OFF : stand-by mode a s S S L TU as S L asN I O E S L eW st R N e s I L E L N I Tlow pwc_cfgL Hop Clk S AS Power Switches Thigh (xN) L N S LS R e s ck Local Clo ck G enera tor Vc ore S e p es eW O U sTt LS LS Vlow cut_off L I P Un it re s e t_ n S L S L S L S L S L S L S L S e s l LS LS V o n e d LS Vhigh ULTRA Vuco CUT-OFF IP units have 5 supply modes S S LP M Vco re I V ig h h I R N Vhigh CONTROL HOPPING CONTROL S S L L MEM N L SL L LS oN A d L S L O U hut SL SL SL SL S S L h oS P ower Swit ch es LS ort TU Local C lock Gener ator Vc or e LS N O esR Tlo w LS T S L L S S S S S S S L L L L L L L I L O U LS S re s e t _n ep LS S esl LS S S V ig h h V o n e d L L L SL SL LS SL SL S SL SL Clo AS S AS L : DC-DC converter S S L L Hopp ing Unit NI S Th i g h (x N) - LPM ck Cl o Local Clock G enerat or Vco r e L S L S L u c t _ fo f reset_n si gnal s ck L dcd c_ctrl si gna ls UC O u c t _of _ cl k 80c51 cut_of f sign als I Optionnal L L O T U E a s t Vl ow V hi gh S NoC perf S L S L V hi gh LS SOFTSOFT SWITCH SWITCH e p IN E a s t IN e W s t S L Vhi gh S L S L S L S L S L Vhi gh Vhi gh LS Pausable Clock, Power Supply Unit e s l o N e d A c t iv t y d e t c ti n o N o rt h U O T N o r th U O T e W s t L S L S S L L O T U o S u th IN o S u th LS Network Interface, Test Wrapper, S L S L S S L L O T U o S u th IN o S u th re s e t_ n UCO V o n e d I U O T e R s L S S S L L R N e s Vhi gh Vhigh L O W U e s T t S L S L LPM cut_off_clk S L L L SL I I S L S L S L C lo Each IP core encapsulated with L L V ig h h S S S L S S S S S S S e s l L L L L L V o n e d L L I U O T e R s S L NI S SL SL LS e t AS S A c t d S Local Clock G enerat or Vco r e LS th LS LS N rto re s e t_ n V ig h h LS LS O U N T ro N I o S tu h L h O U T o S tu h R e s SL SL S SL LS e p N o e d I O U T k S S re s e t _n S S S S S S e s l L L L L L S S V o n e d L S SL SL L LPM FHT2 Cloc SA Local Clock Generat or Vc or e LS SL S L V ig h h LS L L SL SL S Power Adaptive NoC Unit Design k S SL V core LS FHT1 oc Cl SA S ck o Cl Local Clock Gener ator Vco r e DC-DC Co nv. Ta r get a v lu e LP M V core N I OFDM C on rt o l ed S wit ch P ower fi lt er Correct or I P Un it IP Unit LPM SD modu lat or dcdc_c on tr ol Tlo w LS - Hoppin g Un it Vcore re s e t_ n P ower Swit ches c u t_ o f S u ct _ of _ f c lk Tlo w L Th g i h( N x) Hoppin g Un it L - Vl ow Vh igh Power Switches c u t_ o f f LS Vlo w V hi gh UCO u c t _ fo _ f c lk S S S UL T RA Vuco ULTRA CUTCUT - OFF OF F Power S witches Hopping Unit c ut_off Thigh T hig h (xN ) SOF S OFTTSWI TC H SWIT CH Tlow T low pwc_ cfg LS Hop Clk + Synchronous Vcore Vcore A SL L l co o to pr S L SL L V h ig h LS LS L V n od e LS S S LS C S L L LS No S S LS s ou sl eep No de Acti vi ty N I No rth O U T No rt h S LS L Vc ore S n ro ch sy n LS ri ai LS Local Clock Generator LS LS SA dl_cfg re set_n l co S oto pr L C No S L s ou on hr nc Sy LS In D a pu t ta O u D tpu at t a k oc Cl Network Interface Network Link Vcore IP Core pwc_ hl Loc al Power Manager V r ef Synchronous V core S D2 ?A L HO PPING CO NT RO L CO NT RO L LS IP Unit Power Supply Selector Vh igh Vlow Vhigh UCO cut_o ff_clk S L LS d ete cti on O U T Re s I N R es Vhigh LS LS LS LS LS O UT West IN E ast LS UCO ULTRA LS LS LS Vlow CUT-OFF 2007 Power Supply Unit HOPPING UNIT Cut-Off Thigh (xN) SWITCH POWER SWITCHES Hop Clk Tlow + CONTROL D•A Vref Vcore from LPM Unit Local Power Supply Unit offers a safe control of internal power supply depending on pre-defined power modes © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential LS LS Asynchronous Node Vuco SOFT- LS LS LS L evel shifte rs + iso lation cells O UT E ast O U T So ut h I N S out h LS LS Power Supply Unit manages Vcore Two power switches Thigh and Tlow LVT transistors A Hopping Unit UCO clock Vhigh An Ultra Cut-Off Generator LS LS LS LS IN West LS LS LS LS LS Networ k Link Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 19 Vdd hopping vs DFS vs DVFS Consommation totale (mW) 20 Vcore = 1,2 V HIGH 18 DFS 16 Vcore à 1,2 V Fclk variable 14 Vhigh = 1,2 V 1,1 V 12 DVS « parfait » Vcore variable Fclk ajustée en conséquence 10 8 1,0 V 2007 LOW 6 Vlow = 0,9 V 4 2 HOPPING 0,9 V sauts entre les modes HIGH et LOW commandé par un modulateur PWM Vcore = 0,8 V 0 0 50 100 150 200 Fréquence du coeur (MHz) Edith Beigné- June 22th, 2009 Confidential 250 300 350 © CEA 2009. All rights reserved Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 20 UCO vs MTCMOS The proposed UCO generator automatically polarizes the gate of the LVT Power switch to its point of minimum leakage 1,00E-07 25°C 45°C 1,00E-08 65°C 85°C Current (A) 105°C 1,00E-09 125°C 1,00E-10 T 1,00E-11 1,00E-12 1, 20 1,30 1,40 1, 50 1,60 1, 70 1,80 Gate vol tag e (V) Compensates for temperature variation, alleviates corners variations. Mode HIGH RESET 2007 OFF VCORE 1,166 V 1,195 V 1,6 mV I(VCORE) 30,4 mA 691 µA 0,93 µA Mode HIGH RESET OFF VCORE 1,194 V 1,192 V 10,5 mV I(VCORE) 13,9 mA 883 µA 7,8 µA Ioff gain Ioff (pA/µm) 169 743 TRX-OFDM Unit using UCO Ioff gain Ioff (pA/µm) 3000 113 MEM Unit using MTCMOS © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 21 Physical Implementation 3440 STM 65nm technology 9 power domains at the top level 6 frequency domains Each block has its own frequency generator and power suppply Low block complexity DC-DC FHT1 FHT2 TRXOFDM 1200 µm RAM1 RAM5 UCO Unit Level Shifters RAM3 NoC Interface Power Rings (VHIGH, VLOW, VSS, VCORE) MEM NoC Perf RAM7 RAM9 Hopping Unit RAM6 RAM12 RAM13 RAM2 TRX-OFDM Unit (250Kgates) 80c51 1600 µm Power Switches RAM4 RAM8 650x50 µm RAM0 RAM10 RAM11 Power Supply 2007 Unit 3480 © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 22 Power / Variability Commonalities Towards the end of Worst-Case design Worst case design in average too pessimistic (over-constrained design) but optimistic in some cases Need for statistical Design & Tools and variability monitoring Towards self-adaptive systems 2007 AVFS (Adaptive Voltage & Frequency Scaling) Knobs & Sensors Adaptive & robust communication means Global system optimization Towards a power efficient and reliable system reaching an optimal Performance/Yield trade-off © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 23 Conclusion Adaptive dynamic power reduction is possible through voltage scaling : VDD hopping or real DVS Management of multi power domains in a complex SoC GALS systems can lead to automatic frequency scaling Synchronization interfaces Asynchronous implementation Leakage problems due to technology scaling are handled through 2007 Insertion of power switches with UCO An optimum Power/Performance/Yield has to be reached today considering Variability issues © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 24 Main References E. Beigné, F. Clermidy, J. Durupt, H. Lhermet, S. Miermont, Y. Thonnart, T. Tran Xuan, A. Valentian, D. Varreau, P. Vivet, C. Chancel, X. Popon, H. Lebreton, “An Asynchronous Power Aware and Adaptive NoC based Circuit”, IEEE Journal Of Solid State Circuits April 2099 Vol 44. E. Beigne, F. Clermidy, J. Durupt, H. Lhermet, S. Miermont, Y. Thonnart, T. Tran-Xuan, A. Valentian, D. Varreau, P. Vivet. “An Asynchronous Power Aware and Adaptive NoC based Circuit”, Symposium on VLSI circuits, Honolulu, June 2008. 2007 E. Beigné, S. Miermont, A. Valentian, P. Vivet, S. Barasinski, F. Blisson, N. Kohli, S. Kumar, « A Fully Integrated Power Supply Unit for Fine Grain DVFS and Leakage Control Validated on LowVoltage SRAMs”, ESSCIRC’2008, Edinburg, UK, Sept 2008. E. Beigné, F. Clermidy, S. Miermont, P. Vivet, “Dynamic Voltage and Frequency Scaling Architecture for Units Integration within a GALS NoC”, Proceedings of the 2nd IEEE International Symposium on Networks-on-Chip, NOCS’2008, New-Castle, UK, April 2008. © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 25 Innovation for industry 2007 © CEA 2009. All rights reserved Edith Beigné- June 22th, 2009 Confidential Any reproduction in whole or in part on any medium or use of the information contained herein is prohibited without the prior written consent of CEA 26