Leveraging MLC STT-RAM for energy-efficient CNN training
Hengyu Zhao, Jishen Zhao · Proceedings of the International Symposium on Memory Systems · 2018
Graphics Processing Units (GPUs) are extensively used in training of convolutional neural networks (CNNs) due to their promising compute capability. However, GPU memory capacity, bandwidth, and energy are becoming critical system bottlenecks with increasingly larger and deeper training models. This paper proposes an energy-efficient GPU memory management scheme by employing MLC STT-RAM as GPU memory to accommodate the image classification training workloads. We propose a data remapping scheme that exploits the asymmetry access latency and energy across soft and hard bits in MLC STT-RAM cells and the memory access characteristics in image classification training workloads. Furthermore, our design enables (i) energy-efficient memory access by leveraging bit-level similarity in training data and (ii) optimal feature map encoding to compress the contiguous 0s in feature maps. Our design reduces VGG-19 and AlexNet training time, GPU memory access energy and capacity utilization by 76% and 70%, 45% and 40%, 26.9% and 26%, respectively.