THE EFFECTS OF NUMERICAL PRECISION IN SCIENTIFIC APPLICATIONS

[[THE EFFECTS OF NUMERICAL PRECISION IN SCIENTIFIC APPLICATIONS.pdf]]

Take aways

  1. Since neither posit nor 16-bit floating-point formats are supported in commercial general-purpose processors

    Can be one motivation? NVIDIA GPUs seems to be the very few available hardware for half precision.

  2. How to evaluate:

    • use three HPC applications and three ML apps
    • Use 64-bit results as ground truth
    • Run those apps with specific formats
    • Compute the MSE with ground truth