SYSTEMS, APPARATUS, METHODS, AND ARCHITECTURES FOR HETEROGENEOUS PRECISION ACCELERATION OF QUANTIZED NEURAL NETWORKS

摘要

For one embodiment, a hardware accelerator with a heterogeneous-precision architecture for training quantized neural networks is described. In one example, a hardware accelerator for training quantized neural networks comprises a multilevel memory to store data and a software controllable mixed precision array coupled to the memory. The mixed precision array includes an input buffer, detect logic to detect zero value operands, and a plurality of heterogenous precision compute units to perform computations of mixed precision data types for the forward and backward propagation phase of training quantized neural networks.