A case study of scientific computing with limited precision numbers: Applying rescaling and baseline shifting techniques to a spectral Vlasov–Maxwell code

We have developed SpectaVla, a code that solves the multi-dimensional Vlasov–Maxwell equations using the spectral method described in G. L. Delzanno [J. Comp. Phys. 301, 338 (2015)]. This code is designed to take advantage of a chip-scale tensor core optimized for 3D convolution with 8-bit fixed-point numbers. The code is benchmarked against theory on a number of fundamental plasma physics problems. We employ rescaling and baseline shifting techniques to ensure proper precision of the results when using 8-bit fixed-point numbers.