Increasing double precision throughput on NVIDIA Maxwell GPUs.

This paper deals with the impact the architectural changes of modern GPUs have on their use in scientific computing. It particularly focuses on significant drops in the number of double precision functional units in NVIDIA Maxwell architecture. Proposed remedies of the potential negative impact on GPGPU applications that are based on multiple precision arithmetics are discussed. Two new algorithms for fast and precise multiplication and fused multiply add for double precision arithmetics emulation are also presented here. Using these methods, we were able to boost the double precision performance of NVIDIA GTX 980 Ti from 95 GFLOPS up to 286 GFLOPS. The proposed methods are applicable also to other GPUs.