Technical Papers Archive

FitFloat: Read/Write Random-Access Compressed Floating-Point Arrays for GPUs


Authors: Andrew Rodriguez and Martin Burtscher (Texas State University)

Abstract: In many floating-point applications that use GPUs, device memory is a bottleneck limiting the maximum problem size. One solution is to use Unified Virtual Memory, but the limited physical GPU memory likely causes page evictions, incurring potentially large performance penalties. We present FitFloat, a drop-in floating-point array replacement supporting user-specified precision on GPUs with the goal of reducing storage requirements of scientific applications while maximizing performance over unified memory. FitFloat targets applications limited by the GPU's memory capacity, allowing larger problem sizes to be efficiently processed. We evaluate the storage savings and performance improvement of our approach relative to using native arrays on 11 single-precision and three double-precision codes from the HeCBench suite on two different GPUs. In the best case, FitFloat yields a speedup of 739 and provides a geometric-mean speedup of 16.1 for the float codes and 20.1 for the double codes on an RTX 4090.



Back to Technical Papers Archive Listing