ABSTRACT
Many DTM schemes rely heavily on the accurate knowledge
of the chip’s dynamic thermal state to make optimal performance/
temperature trade-off decisions.
1. INTRODUCTION
Most of today’s high performance multi-core processors
suffer from the heavy power/thermal stress
(page 1 col 2)
If the statistical characteristics of the power dissipation
profile does not change in time, Kalman filter based approach
can generate optimal thermal estimates using sensor observations
[4].
In this paper, we investigate the problem of
adaptive temperature tracking at runtime by considering the
dynamic changes in the statistical characteristics of the power
profile.
2. PRELIMINARY
2.1 System Dynamics
2.2 Kalman Filter Based Thermal Tracking
3. PROBLEM DEFINITION AND
CHALLENGES
The statistical characteristics of each potential power state
could be captured by simulating or experimenting with all
potential applications sets (integer vs floating point, scientific
vs multimedia and etc.)
4. ADAPTIVE TRACKING BASED ON
RESIDUAL WHITENING
4.1 Autonomous Detection
In this section we explain how Kalman filter can be used
to autonomously detect the switch of the power states.
Note that we use Cs instead of C[n|n−1] since we assume
the system has reached the steady state.
Once again we use steady state Ks as parameter.
(p4)
Basically, we evaluate error between observation and prediction
and estimate the autocorrelation.
