An accelerator for posit arithmetic targeting posit level 1 bLas routines and pair-HMM

Abstract

The newly proposed posit number format uses a significantly different approach to represent floating point numbers. This paper introduces a framework for posit arithmetic in reconfigurable logic that maintains full precision in intermediate results. We present the design and implementation of a L1 BLAS arithmetic accelerator on posit vectors leveraging Apache Arrow. For a vector dot product with an input vector length of 106 elements, a hardware speedup of approximately 104 is achieved as compared to posit software emulation. For 32-bit numbers, the decimal accuracy of the posit dot product results improve by one decimal of accuracy on average compared to a software implementation, and two extra decimals compared to the IEEE754 format. We also present a posit-based implementation of pair-HMM. In this case, the hardware speedup vs. a posit-based software implementation ranges from 105 to 106. With appropriate initial scaling constants, accuracy improves on an implementation based on IEEE 754.