Binary and ternary convolutional neural network acceleration by in-nonvolatile memory computing with Voltage Control Spintronics Memory (VoCSM)

We report a novel convolutional neural network (CNN) accelerator utilizing “voltage control spintronics memory” (VoCSM). High throughput processing is achieved by high speed in-“nonvolatile memory”-computation using high density VoCSM array. Since VoCSM has largest endurance even in short write pulse of all nonvolatile memories, write access speed can be increased. Also, density of processing element is higher than conventional SRAM based one, throughput is further increased. These technique reduces latency of CNN by 46% (ternary) and 73% (binary) compared to conventional CMOS processor with 8bit integer (INT8) for CIFAR-10 classification task. Also, energy is reduced by 72% (ternary) and 86% (binary).