Variability-Aware Thermal Simulation using CNNs

With rising power densities in modern-day electronic systems, temperature has emerged as a fundamental design constraint. This has led to the advent of a range of thermal-aware design and runtime management techniques. However, such techniques are heavily dependent on a fast and accurate thermal modeling method. These methods need to account for manufacturing variability, that significantly impacts the chip's power and performance. Similarly, leakage power too contributes to a substantial portion of the total power. Thus a thermal modeling method can be accurate only if it is capable of incorporating the effects of process variation as well as leakage power. In this paper, we propose a simple and elegant residual convolutional neural network for thermal estimation in the presence of variability, which leverages the physics of heat transfer. Our approach is capable of modeling modern-day 3D chips with microchannels and incorporates accurate leakage power models. To enable ultra-fast thermal estimation, we implement our technique on a GPU. Our experiments show that our technique is orders of magnitude faster than the state-of-the-art with a similar, if not better, accuracy. The mean absolute error using our technique is 0.61°C, for a maximum temperature rise of 67.5°C (0.9%).