A 40-nm 13.88-TOPS/W FC-DNN Engine for 16-bit Intelligent Audio Processing Featuring Weight-Sharing and Approximate Computing

This work presents a novel accelerating engine to efficiently compute the fully connected (FC) deep neural networks (DNN) based on weight sharing (WS). Beyond the highly reduced weights and weight accesses in previous designs, the proposed engine significantly reduces the multiplications by exploiting the distributive law of shared weights. All-digital in-memory computing (IMC) and approximate computing (AC) were applied to accelerate accumulations and improve energy efficiency. We have designed and implemented a test chip in 40nm CMOS for intelligent audio processing of dysarthric voice conversion (i.e., using an FCDNN composed of neurons with 16-bit inputs, 16-bit weights, and 16-bit activations). When turning AC off and on, the energy efficiency for 16b operation at 0.9V is 7.04 TOPS/W and 13.88 TOPS/W, respectively.