A Reconfigurable Fused Multiply-Accumulate For Miscellaneous Operators in Deep Neural Network

In this paper, a novel general-purpose multiple-precision reconfigurable fused multiply-accumulate (RE-FMA) is proposed for deep neural network. Data pre-processing, post-processing, look-up-table (LUT) are added, and accumulate functions are enriched. In addition, the various tensor operations are performed individually, requiring only one data buffer. The tensor operations include matrix multiplication, softmax function, dot-product, quantized-add, max-pooling, average-pooling, etc. The RE-FMA flexibly combines the basic elements to accomplish the tensor operators with a smaller area compared to traditional FMAs. It also features high speed and energy efficiency. The RE-FMA has significant potential to accelerate neural network inference in resource-constrained devices. Experiment results show the RE-FMA with parallelism of N = 32 is able to operate at a maximum frequency of 625 MHz. The full-load power consumption and area are only 131.92 mW and 0.126 mm². Compared to the sum of four dedicated operator hardware accelerators, the reductions are 19 % and 25.6 %, respectively.