An RRAM-Based 40.6 TOPS/W Energy-Efficient AI Inference Accelerator with Quad Neuromorphic-Processor-Unit for Highly Contrast Recognition

We present a non-volatile edge deep neural network accelerator with a resistive-gate FinFET (RG-FinFET) memory and a parallel processor for edge AI inference. The RG-FinFET has the potential for 8-level operation. In the system, data storage and multiplication are carried out in the RG-FinFET array, and all the other operations are performed in a 4-core neuromorphic processing units (NPU). Quantization error is introduced into training stage through ex-Situ quantized training method, thus, the accuracy can still reach 97.24% and 80.18% respectively for MNIST and CIFAR-10 datasets while the parameter capacity is nearly 8x smaller. Eventually, the system's computation efficiency with 40.6 TOPS/w can be achieved, which is well-suited for the end-to-end integer-only AI-Inference hardware in CIM.