ISSCC 2002 / SESSION 25 / PROCESSOR BUILDING BLOCKS / 25.7 25.7 Adaptive Body Bias for Reducing Impacts of Die-to-Die and Within-Die Parameter Variations on Microprocessor Frequency and Leakage James Tschanz, James Kao', Siva Narendra, Raj Nair, Dimitri Antoniadis', Anantha Chandrakasan', Vivek De Microprocessor Research Labs, Intel Corporation, Hillsboro, OR, 'Massachusetts Institute of Technology, Cambridge, MA Measurements on a 150nm CMOS test chip show that on-chip bidirectional adaptive body biasing compensates effectively for die-to-die parameter variation to meet both frequency and leakage requirements. An enhancement of this technique to correct for within-die variations triples the accepted die count in the highest frequency bin. See Digest page 422 J. Tschanz Block Diagram Outline _____________________________ -. Process parameter variations I I 1 - Die-to-die (D2D) - Within-die (WID) Adaptive body bias (ABB) test chip Effectiveness of ABB Impact of within-die variations Compensation of WID variations (WID-ABB) 1ll; ,5 I I I 4 L I I I/ ob I -6- I - ; Phase detector & counter Cntzal path , II A I li 'i!! Parameter Variation Phase Detector and Counter I__ I 150nm CMOS technology PD /I e '1 Critical Dath - , I 0.925 1 1.075 1.15 I out 1.225 Technology (one subsitel Dimensions I I 150nm CMOS I 1.6X0.2 mm2 I -0.5V to 0.5V Body bias resolution ----= 344 -- - - = ---e-=-- 32mV ---__ ______ & - _ _ _ ISSCC 2002 VISUALS SUPPLEMENT / 0 IEEE _ _ I I code \ ! PD=I: F> , F, PD=O: ,F , < F, 4.5X6.7 mm2 + Bias range 5-bit digital "B"t "ss Adaptive Body Bias (ABB) U1 I RBB i j, Body Bias Modes - I . NBB + FBB 1 'CCA 'REF FBB + RBB NBB 4 RBB '= vCCA vCC 'CCA = 'CC 'CCA VREF< VCCA Count: PD=O Accounts for WID RBB < 'CC VREF< VCCA Count: PD=I Count: PD=I Area overhead: 4 5 3 % ABB Test Chip Operation 1.2 , 1.8 I Frequency target 1.5 1.2 U -8 .- s 0.9 0.8 al -p c 0.6 0.7 L I 0.5 0 20 9 0.3 ' 0 0.6 40 80 60 Time (ms) lj/i Body bias adapts to meet frequency target with minimum leakage Neglects WID variation I bias (miniyize P), I 100% E 80% 60% 40% 6 20% 0% nl 115 1225 IIN:RBB, N:FBB N: FBB I . I I t 1 die 1 Effectiveness of S-ABB B 1075 I Area overhead: -2% [paper 16.41 r 0.4 -04 H' ______ and critical path 1 Normalized frequency s Measure Pleak of circuit block p~= Phase detector 0925 10 dies 13dies 36 dies Frequency vs. Critical Path Count (NCP) 11oc Accepted dies Frequency Variation $ 5 % - 4 - 3 =8 2 09 1.I 1.5 13 Normalized die frequency E z 2 1 / Z i 0925 - 1 1075 115 1225 Normalized frequency - _=== ==r-* _ I I ~ I = - _ c l _=__- I .____--_- =- Jl- I Frequency p and CJ reduce as N,, increases Frequency distribution unchanged for N , > 14 - _I___ - --= - --- Continued on Page 538 0 IEEE / ISSCC 2002 VISUALS SUPPLEMENT 345 ISSCC 2002 / PAPER 25.7 ISSCC 2002 / PAPER 25.6 -_ Continued from page 345 _ _ _ _ -- ontinued from page 343 I__ I _ V _ h - I Middle Bwass - __- _I- _ l _ _ _ l _ _I_-&- ___-l__l__- WID Delav Variation vs. Logic Depth /I 40% OS olp = 5.6% E 20% (0 os olp = 3.0 Y o E I RFread.ebp, -16% -8% 0% 8% 16% Variation (%) - Active low inputs and static receivers reduce impact of noise 4.27% Device olp need fnr a lntrh -- __ - - _ I Path Depth dynamic nodes to allow phase 1 (CK)inputs to be "caught" removing the [ Late Bypass pseudo-NMOS Mux selects enable pullup and latch reducing power and enabling stall recovery Within-Die Adaptive Body Bias (WID-ABB) Ji Compensates for WID variation /I j/ : N __________ e _*_____=-I , _ _ - -__- I _ _^- Results Datapath ? $ g a * Functional units (FU) are pre-enabled saving power and removing the need for a result mux Effectiveness of WID-ABB - /; Z 100% 80% 60% 40% 20% 0% ; FU outputs flow transparently through the middle and late bypasses stopping only for the CK clock boundry at the input to the FU Accepted 97% in ahighest WID-ABB I 1 FEEKyI % 5 x m - bin 4 I= 23 Antimillers are used to enable long precharge pulldown results busses E z 1 I I 0 0925 1 1075 115 1225 Normalized frequency Summary Ii Within-Die Bias Distributions _-- A 7.48mm2 integer datapath i s constructed with 128 general registers,6 fully bypassed ALUs a n d 4 cache address ports - Double-pumped decoders/wordlinesin register file delivered high performance/bandwidthin a small foot print 96 GB/s read bandwidth, 64 GB/s write bandwidth in 2mm2 - Antimiller circuits reduce area by allowing long minimally spaced precharge-pulldownnets - Register file write bitlines are reused to accomplish first stage of bypass - Pre-enabledfunctional units reduce power consumption 15% while removing logic stages from bypass network - NMOS zody 0.5 RBB Bias(V) WFBB _ _ _ . a - 538 ISSCC 2002 VISUALS SUPPLEMENT / 0 IEEE __.__---- ISSCC 2002 / PAPER 26.1 1 Bias Resolution 4' I dies, 100% 100 % I 1' 1 i It I B I i 1' ! 1 WID -ABB dies, F 1.075 2 Yo 1.89 % 0.50%1 1.47%1 66% 0.69 Yo 97 % 0.21 % Distribution of White Noise ('ds) packaged parts measured on an automatic handler n 1 hlwi: 0.035 SD: 0.00s T(C): -40 i 300mV bias resolution sufficient for ABB WID-ABB requires 1OOmV bias resolution Conclusion D2D and WID variations impact microprocessor frequency and leakage ABB improves die acceptance rate from 50% to 100% ABB is most effective when WID variations are considered Compensating for WID variations by WIDABB increases number of high frequency dies from 32% to 97% f 1 1 F 6.02 4 d.04 d.06 Theorbtical BrownianNoise Conclusion The measured angular rate white noise is 0.05 deg/rt.s with an ultimate resolution of 50 deg/h. 'I The gyros operate accurately -4OC to +85C during applied shocks over 1000 gees and survive shocks of 30,000 gees unimpaired. i 1 Complete angular rate to voltage transducer in 7mm x 7mm x 3mm package weighing 0.35 grams. ! 0 IEEE / ISSCC 2002 VISUALS SUPPLEMENT 539