Vt Variation Effects on Lifetime eliability Smruti. Sarangi Josep Torrellas University of Illinois at Urbana-Champaign
Motivation Chips are wearing out faster as technology scales >>180nm: 20 year design lifetime 130nm: 10 years 65nm: 7 years 32nm:??? How does process variation affect this trend? 2
Vt Variation andom Within-Die 3
FPQ FPMap FPMul FPAdd FPeg BPred Vt Variation UL2cache1 L1D IntMap UL2cache2 IntQ L1I IntExec Integ DTB ITB LdStQ UL2cache3 Systematic Within-Die 4
Vt Variation Die-to-Die 5
Vt Variation Component Variation σ/µ andom WID 5.2% Systematic WID 5.2% Die-to-Die 5.2% Total Vt variation σ/µ = 9% 6
Time To Failure Probability TTF 1% MTTF Time to Failure 7
Failure Mechanisms TTF depends mostly on temperature Exponential temperature dependence Time-Dependent Dielectric Breakdown (TDDB) Electromigration (EM) Stress Migration (SM) Quadratic temperature dependence Thermal Cycling (TC) 8
!""# $%&'() *+,-%&./(0 1+ 12( $%&'() %1 34 56 "2(!"#$%&'(!""# $%&'() *+,-%&./(0 1+ 12( $%&'() 56 "2(!"#$%&'( "+ &.*( )2+7) 12( 8+-9.*(0 (::(81 +: %1 %&&34:+', :%.&',( -(82%*.)-)6 Lifetime vs Temperature &.*( )2+7) 12( 8+-9.*(0 (::(81 +: %&& :+', :%.&',( -(82%*.)-)6 Normalized MTTF Normalized MTTF 10 1 10 Combined Combined TC TC TDDB TDDB SM SM EM EM *%-. 72.&( ')( 12 :,+)'CC& C,+C+ 1 W* T 1. 65 70 75 80 85 90 95 100 105 60 65 70 75 80 85 90 95 100 105?T0 @A Temperature (C) Temperature (C) 12( 12 #.;6 <6 =+,-%&./(0!""# :+, "2(,-%& 5>8&.*;?"5@A ".-( B(C(*0(*1 8+*)1 #.;6 <6 D,(%E0+7* =+,-%&./(0!""# :+,Greskamp "2(,-%&?F!@A 5>8&.*; ".-( B(C(*0(*1 9 Brian B.(&(81,.8?"BBD@A F1,())!.;,%1.+* %*0?"5@A G&(81,+-.;,%1.+* [L*0.1 0.1 60
Lifetime vs Temperature Microprocessor MTTF (180nm SOI) 100,000 MTTF (Years) 10,000 1,000 100 55 65 75 85 95 105 115 Junction Temperature (C) 1.3 V 1.2 V 1.4 V Image: Freescale Semiconductor 10
0 105 hermal Cycling Migration (SM), ould change re dominant. tems; if any capture this 0 small cells. ognormally- sa Heat spreader Closing the Loop Vt cs2 Vteff Ps Cell 1 T cs1 MTTF Die surface Fig. 2. Thermal model and electrical equivalent of processor die, heat spreader, and sink. dependence = heat Linear = Exponential dependence and static power are both proportional to:! "2! " Vtef f c2 kt Ps µ exp q c3 kt /q Vtef f = V t c1 (T T 0) The mobility µ itself has a temperature dependence; we c /T 4 M T T 65nm F emodel has c1 = 5.0 10 4, use µ T 1.5. HotSpot s c2 = 3.9 10 2, and c3 = 1.3. With these constants, the leakage model predicts a 41% increase in static power consumption as die temperature changes from 75 C to 90 C, even 11
0 105 hermal Cycling Migration (SM), ould change re dominant. tems; if any capture this 0 small cells. ognormally- sa Heat spreader Closing the Loop Vt cs2 Vteff Ps Cell 1 T cs1 MTTF Die surface Fig. 2. Thermal model and electrical equivalent of processor die, heat spreader, and sink. dependence = heat Linear = Exponential dependence and static power are both proportional to:! "2! " Vtef f c2 kt Ps µ exp q c3 kt /q Vtef f = V t c1 (T T 0) The mobility µ itself has a temperature dependence; we c /T 4 M T T 65nm F emodel has c1 = 5.0 10 4, use µ T 1.5. HotSpot s c2 = 3.9 10 2, and c3 = 1.3. With these constants, the leakage model predicts a 41% increase in static power consumption as die temperature changes from 75 C to 90 C, even 12
0 105 hermal Cycling Migration (SM), ould change re dominant. tems; if any capture this 0 small cells. ognormally- sa Heat spreader Closing the Loop Vt cs2 Vteff Ps Cell 1 T cs1 MTTF Die surface Fig. 2. Thermal model and electrical equivalent of processor die, heat spreader, and sink. dependence = heat Linear = Exponential dependence and static power are both proportional to:! "2! " Vtef f c2 kt Ps µ exp q c3 kt /q Vtef f = V t c1 (T T 0) The mobility µ itself has a temperature dependence; we c /T 4 M T T 65nm F emodel has c1 = 5.0 10 4, use µ T 1.5. HotSpot s c2 = 3.9 10 2, and c3 = 1.3. With these constants, the leakage model predicts a 41% increase in static power consumption as die temperature changes from 75 C to 90 C, even 13
)'9)1,%1(,().)1%*8( csi 1+,(%82 12( 2(%1 )C,(%0(,6 #.*%&&>A 12( 2(%1).*E 0.)).C%1() 2(%1 12,+';2 12(,-%&,().)1%*8( se 1+ 12( (*$.,+*-(*1A 72.82.) %))'-(0 1+ 9( %1 NL 56 S( %))'-( )1(%0>)1%1( +C(,%1.+*A 12(,(:+,(.;*+,.*; 2(%1 )1+,%;(.* 12( 12(,-%& 8+-C+*(*1)6 T00.1.+*%&&>A %) )2+7*.* #.;',( PA 7( %&&+7 &%1(,%& 8+*0'81.+*.* 12( 0.( 9> -+0(&.*; % &%,;( 12(,-%&,().)1%*8( 9(17((* %0U%8(*1 8(&&)6 Temperature Model Environment (45 C) Heatsink se Heat spreader cs2 cs1 Cell 1 Heatsink Heat spreader Package Die Die surface #.;6 P6 "2(,-%& %))(-9&>?,.;21@ %*0 (&(81,.8%& (V'.$%&(*1?&(:1@ +: C,+8())+, 0.(A 2(%1 )C,(%0(,A %*0 2(%1).*E6 Modeled with HotSpot (Skadron et al.) 0(*8( %*0 +*&> % :'* 12( FC2(, +61_) M<PO 8(&&) )(C%,(&%1(06 #+ 12( &(*;12 T:1(, 12 -.*(0A 7( $%,.%1.+* 8 0.() 7.12 1,%*).)1+, "2.) 8+-C %*0 *+ )C D%)(0 σvt0 = 0.0 8+*1,.9'1.+ 7( )(1 (0.09/ 3 14
Experiment For i = 1 to 20000 do 1. andomly generate die i Vt variation map 2. Partition die into 1000 equally-sized cells 3. Compute temperature for each cell 4. Generate lifetime distribution for each cell 5. Sample lifetime distribution for each cell 6. Die i lifetime = min(cell_lifetimes) 15
Processor Model Model Intel Core Solo 65nm floorplan Per-unit dynamic power modeling SESC cycle-accurate simulator with WATTCH Profile from crafty SPECint benchmark Steady-state dynamic power = 14W HotSpot thermal model 16
Example Die Before Variation DTB BPred Integ IntMap FPMap FPQ IntMap FPMul IntQ FPQ IntQ FPAdd FPeg BPred ITB LdStQ ITB LdStQ FPAdd FPeg FPMul FPMap L1D L1I L1D L1I MTTF=0.90 DTB IntExec IntExec Integ After Variation MTTF=0.75 MTTF=0.95 MTTF=0.85 MTTF=1.0 UL2cache1 UL2cache2 UL2cache3 UL2cache1 UL2cache2 UL2cache3 MTTF=0.8 Fig. 6. Spatial distribution Brian of cell Greskamp MTTFs before and after variation for an example die. Color and contours indicate MTTF. extreme va temperature than norma C. Future T Assumin on lifetime 17 scales, per-
65< se G 3&, %!,!%!P# $.# *#6!&(!6!$)!%A&3$ 5H Vt0 L&*!&$!5, esult: Vt Variation $ &6"5 ".5<" $.&$ *#6&$!L#6) "%&66 3.&,-#"!, se 3&,.&L# Impact on <.#, Lifetime $*5,- *#6!&(!6!$) 35,"#]+#,3#" Pleak0!".!-.? @NVS1 U ^ 1FM2@U_` U` T T F1% O_^ FUOO1^1`@ ansm1/ _O J1N@/U`b @J1^QNS ^1/U/@N`21 N`F in _O TTF `_ 7 an^un@u_` 0_=1^? eduction Vt Variation 1% due to S1NbNc1 Pleak0 (W ) X?K 9I?I 9T?K 9K?I I?d 7T[ 7T[ 7W[ 79>[ se (K/W ) I?X I?W I?e 7:[ 7>[ 7>[ 7d[ 799[ 79e[ 79T[ 7TK[ 7KI[ 7:X[ 7d9[ 7dX[ Ua? 2 _`2SM/U_`/ Brian Greskamp 9?I 7e[ 7:>[ 7dT[ 7d>[ 18
Conclusion D2D and systematic variations are biggest problem for aging Potential solutions educe overall leakage Increase heatsink size 19
Vt Variation Effects on Lifetime eliability Smruti. Sarangi Josep Torrellas University of Illinois at Urbana-Champaign