An Energy-Efficient Gain-Cell Embedded DRAM Design with Weight Encoding for CNN Applications.

In the inference process of convolutional neural networks (CNNs), reading weights from on-chip memory numerous times consumes a significant amount of energy, which is one of the bottlenecks for low-power design. Based on the analysis of the characteristics of CNN weight distribution and the reading mechanism of GC-eDRAM, an energy-efficient GC-eDRAM design with weight encoding is proposed for CNN inferences. By minimizing the instances of reading bit ‘1’, the proposed GC-eDRAM design achieves a reduction of energy dissipation in reading operations for CNNs, as indicated by the post-simulation results, with a decrease of 13.7% to 16.4%.