ECE Dept., University of Patras
A Simple Mechanism to Adapt Leakage-Control Policies to Temperature
Stefanos Kaxiras
Polychronis Xekalakis
Georgios Keramidas
[email protected] {kaxiras,keramidas}@ee.upatras.gr
ECE Dept., University of Patras
Introduction z
Substantial rise in heat density
z
Increasing need for temperature aware micro architectures
z
Deep sub-micron/nano era: leakage control necessary
z
Leakage greatly depends on temperature Î We propose temperature-adaptive leakage control policies for on-chip caches
ECE Dept., University of Patras
Outline z
Why Thermally Adaptive Leakage Control Techniques?
z
Cache Decay, Drowsy and Hybrid Schemes
z
Our Proposal: Thermally Aware Hybrid Mech.
z
Temperature Sensitive Timer: 4T DRAM Cell
z
Evaluation of Hybrid vs. Decay and Drowsy
z
Summary - Future Work
ECE Dept., University of Patras
Why do we Need Thermally Adaptive Leakage-Saving Techniques? z
Static consumption becomes substantial • Leakage control • Caches primary target: They account for a large fraction of the ‘T’ budget
z
Some mechanisms trade static for dynamic
z
Dominant leakage contributor (subthreshold leakage) depends exponentially on temperature
z
“Stationary” static/dynamic trade-off is clearly not optimal
ECE Dept., University of Patras
Main Idea z
Gated approach (cache decay) to save leakage: • Saves lots of leakage, destroys data • Dynamic power penalty if wrong (miss)
z
Tradeoff (save leakage but incur misses) changes with temperature! • High T (high leakage) Î — aggressive gating, don’t mind misses
• Low T (low leakage) Î — very careful gating, misses hurt — also use DVS!
ECE Dept., University of Patras
Cache Decay and Drowsy Cache Cache Decay: wait a “decay interval” after the last access and gate the cache line z Decay interval defines behavior: • “decay” the cache lines soon Î increase in dynamic power • “decay” the cache lines late Î might not save enough leakage z Drowsy Cache: DVS in cache lines • leakage savings smaller compared to decay • but: no dynamic power penalty z Hybrid approach captures the benefits of both z
ECE Dept., University of Patras
Hybrid Mechanism If line idle for a few cycles we put it in drowsy mode z If a drowsy line remains idle for a longer period we decay it
z
Miss
y Accesses Enter Drowsy Mode
Leak
Leak less
Last access: line idle until replacement (Dead Time)
Drowsy exploits DI z But: Is this enough? z
Perform Decay
No leak Timeline
ECE Dept., University of Patras
Hybrid Thermal Aware Mechanism z
We keep the drowsy interval const & vary the decay interval according to temperature Line Accessed
Perform Drowsy
Perform Decay
Miss
Leak Line Accessed
Perform Drowsy
Leak less Perform Decay
No leak Leakage Power Saved Over simple Hybrid ~ 24%
Miss
Leak
less
35ο C low-leak
No leak
85ο C high-leak
ECE Dept., University of Patras
Timing Mechanism Supply Regulator Tick every 512c
2 state FSM
Refresh upon line access
4T
Power Line Cache Line 6T
Low Leak Inverter
Drowsy timer: 2-state FSM, reset on access, ticked by global timer (512c) • Cache line idle 512c Î drowsy mode z Decay timer: temperature-sensitive 4T cell, • charged on access • “decays” after some temperature-depended decayinterval Î gates cache line z
ECE Dept., University of Patras
4T Cell’s Timing Nature 140000
Decay Cell Retention
100000
Decay cell with Different z Gate Transistor
80000 60000 40000 20000 0
35
45
65
85
Temperature (Celsius)
4T cell’s retention time function of temperature • Previously proposed as cheap temperature sensors
120000
.
Retention Time es (cycles)
z
Can “design” retention times • gate transistors, geometry z Automatic adaptation of retention times with T z Retention times 110 converge to the same value This is what we want!
ECE Dept., University of Patras
Hybrid-Decay-Drowsy, Which is Better? z
HotLeakage, 6 Spec, PSpice
z
Overall savings depend on: • L2 latency Î increases relative energy delay —
in latency favors drowsy
• Tox Î reduces leakage —
in Tox favors decay
z
If decay is better than drowsy hybrid is by default better than both
z
Parameters chosen to favor drowsy
ECE Dept., University of Patras
Evaluation: Decay & Drowsy vs. Temp. 80% 70%
Normalized Leakage =
60%
35ο C
Decay Drowsy
45ο C
35ο C 45ο C 65ο C 85ο C 110ο C
50%
New Leak + Dyn. Ovh. Old Leakage
40% 30% 20% 10% 0%
65ο C 85ο C 110ο C
800
1600
3200 6400
12800 25600
Decay Interval
51200 102400
z
Drowsy Î flat, does not depend on T nor DI
z
Decay Î depends leak/dynamic ratio, affected by T (leakage) and DI (dynamic)
ECE Dept., University of Patras
Evaluation: Hybrid vs. Drowsy 65ο C
45ο C
35ο C
19%
Normalized Leakage =
Hybrid Drowsy
ο 17% 85 C
15%
ο New Leak + Dyn. Ovh. 13% 110 C
Old Leakage
11% 9% 7% 5%
Hybrid 800
1600
3200
6400
12800
25600 51200 102400
Decay Interval
z
Hybrid easily outperforms drowsy
z
Hybrid adaptivity: simply chose the “best” decay intervals @ each temperature
ECE Dept., University of Patras
Approximation of ideal hybrid with 4T 15%
Drowsy
13% Normalized Leakage = 11% New Leak + Dyn. Ovh. 9% Old Leakage 7%
Adaptive hybrid using “best” decay intervals
5% 35
45
65
Temperature (Celsius)
85
110
ECE Dept., University of Patras
Approximation of ideal hybrid with 4T 15%
Drowsy
13% Normalized Leakage = 11% New Leak + Dyn. Ovh. 9% Old Leakage 7%
Adaptive hybrid using “best” decay intervals
Adaptive-hybrid using 4T decay intervals
5% 35
45
65
Temperature (Celsius)
85
110
z
We approximate the ideal hybrid with the 4T decay mechanism
z
Easily superior to drowsy, especially for high temps
ECE Dept., University of Patras
Summary z
Hybrid Temperature Adaptive Leakage Control: • decay+drowsy • decay intervals vary upon temperature • High temperatures: aggressive decay • Low temperatures: employ drowsy much more, to keep the dynamic power low
z
4T decaying cells as timers to control decay intervals
z
Outperform best of non-temperature adaptive schemes by as much as 33% @ high T
ECE Dept., University of Patras
Future Work z
Thermal DVS
z
Thermally Dynamic Threshold Transistors
z
Varying Drowsy also, under some constraints may produce good results
z
Reducing power means reducing Temperature Î quantify the temperature reduction achieved by hybrid
z
Model impact of sensor errors