Thread: Towards fine-grained precision reconfiguration in variable-precision neural network accelerator

Abstract

In this work, we propose a neural network accelerator that supports finer-grained precision tuning for both activations and weights. To the best of our knowledge, this is the first neural network accelerator supporting arbitrary bit widths within 8 bits for both neural weights and activations. Our accelerator combines the features of bit-parallel and bit-serial design methods so that we can accomplish high arithmetic-unit utilization and precision tuning flexibility simultaneously. According to the cycle accurate simulation and the synthesized result, the proposed accelerator achieves up to 1.77x energy-efficiency improvement over the bit-serial design of Stripes for the evaluated workloads.