A Variable and Extended Precision (VRP) Accelerator and its 22 nm SoC Implementation

High-Performance Computing (HPC) scientific applications require linear solvers and eigensolvers, and the growth in the size of linear systems has led to the broad use of iterative projection methods, which have lower memory occupancy than direct methods, but suffer from roundoff errors. Augmenting the precision speeds up convergence, but currently only software libraries support variable and extended precision Floating Point (FP) computation beyond 80 bits.We present the VaRiable and extended Precision Accelerator (VRP), a RISC-V accelerator, implemented on a System-on-Chip (SoC) using GF22FDX technology. It supports FP computations with a number of significand bits ranging from 2 to 512. The VRP provides an average 19.25x application speedup compared to the well-known MPFR software library running on a 2400 MHz Intel Xeon processor. Furthermore, extended precision enables the convergence of linear solvers for problems that would not otherwise converge.