Low-Precision Vectorized Arithmetic Unit Designs for Deep Learning

To meet the demands of low-precision operations in deep learning, this paper presents two architectures for low-precision computation units. Firstly, we propose a low-cost and multi-precision multiplication unit to satisfy various precision requirements. Secondly, to satisfy the requirements of high-performance floating-point operations, we investigate the implementation of floating-point fused dot-product units and devise a scalable architecture supporting low-precision mixed operations. Experimental results demonstrate that compared to conventional multiplication units with the same function, the proposed multi-precision multiplication unit reduces area by 26.6 %; The dot-product units realized by the proposed dot-product architecture reduce the area by over half compared to equivalent arithmetic units, and the proposed dot-product architecture can easily implement fused dot-product units with higher throughput. The amalgamation of these two studies will enhance the floating-point computing capabilities significantly for deep-learning applications.