Emerging Floating-Point Formats for HPC

Floating-point operations are indispensable for many scientific applications. Their precision formats can significantly impact the power, energy consumption, memory footprint, performance, and accuracy of applications. Moving towards exascale, optimizing precision formats in HPC scientific applications could address some key challenges identified on exascale systems. Posit an alternative to IEEE 754 floating-point format, has gained increasing attention in the HPC community because its tapered precision can achieve higher precision than IEEE 754 format using the same number of bits. Still, its precision improvements in general HPC scientific applications require systematic efforts to understand and quantify, which motivates our study in this paper. Our work provides for the first time a Posit implementation of the popular NAS Parallel Benchmark (NPB) suite. Our Posit implementation exhibit 0.4 to 1.6 decimal digit precision improvement in all tested NPB kernels and proxy-applications compared to the baseline. In addition, we show that Posit could benefit a broad range of HPC applications but requires low-overhead hardware implementation.

Reference Publication: Posit NPB: Assessing the Precision Improvement in HPC Scientific Applications, Chien, Steven W. D. and Peng, Ivy B. and Markidis, Stefano, Parallel Processing and Applied Mathematics, 2020, 301โ€”310, Springer International Publishing

Figure: Binary formatting of IEEE Float and 32 bit Posit with es = 2 when representing