Burnt Sienna Team 2022

Xuesi Chen, Daniel Ernst, Margret Riegert

Hotspot Mitigation for Systolic Array CNN Accelerators

On-chip thermal hotspots have become an increasing concern in recent generations of ML accelerators. With the increasing demand for neural network processing from ML applications such as natural language processing or image recognition, and deeper and larger emerging neural networks, a software tool that can accurately predict the appearance of thermal hot-spots has the potential to improve thermal performance of future ML accelerators. Our project, NNShim, builds upon the existing systolic ML simulator SCALE-Sim for data movement and the thermal hotspot simulator HotGauge to generate temperature and power maps, with a specific neural network model and hardware configuration. Our simulation results prove that decreasing clock frequency, increasing aspect ratios and applying row and column skipping are effective ways to cool ML accelerators and mitigating hot-spots.

Related Tech Notes

Project Poster