FastSpot: Host-Compiled Thermal Estimation For Early Design Space Exploration

Darshan Gandhi,  Andreas Gerstlauer,  Lizy John
The University of Texas at Austin


Abstract

Power and temperature of modern day systems have become important metrics in addition to performance. Static and dynamic power dissipation leads to an increase in temperature, which creates cooling and packaging issues. Furthermore, the transient thermal profile determines temperature gradients, hotspots and thermal cycles, which influence wear-out and reliability. Fast and accurate methodologies to facilitate estimation of temperature are needed during design space exploration. Traditional solutions rely on cycle-accurate simulations of detailed micro-architectural structures and are slow. We propose an approach that integrates accurate thermal estimation into existing host-compiled simulations. The developed methodology can incorporate different thermal models, such as HotSpot or discrete-time temperature evaluation models (DTTEMs). Using DTTEM as our thermal model, we show a 32000x increase in simulation throughput for temperature trace generation, incurring an average 0.06K error in transient trace generation and 0.014K error in steady state temperature measurements when compared to a cycle-accurate reference method. A comparative study between HotSpot and DTTEM based on speed/accuracy tradeoffs is reported for the thermal estimation process, showing that a 9x increase in simulation throughput can be achieved using DTTEM as the thermal model while incurring only marginal temperature measurement errors.