Dynamic Thermal-Predicted Workload Movement with Three-Dimensional DRAM-RRAM Hybrid Memories for Convolutional Neural Network Applications

Abstract

Nowadays, Convolutional Neural Network (CNN) is widely used in many applications. Multi -layered convolutional neural networks need lots of memory capacity and bandwidth. A large number of the CNN parameters cause long latency for the memory accesses. To solve this problem, the 3D stacked DRAM-RRAM hybrid memory is discussed. However, the 3D stacked DRAM-RRAM hybrid memory may result in serious thermal problem for the thermal limitation of the DRAM and RRAM chips. In this work, we propose the dynamic thermal-predicted workload movement (DTPWM) to solve this problem. If the overheated banks of the DRAM and RRAM chips are predicted, DTPWM can move the workloads to other non-overheated memory banks. In our experiment, the latencies of the 3D stacked DRAM-RRAM hybrid memory is reduced by 27.7% under the thermal limitation.