Comment by 37ef_ced3

16-bit floating point (as a weight storage format, not used for math) is essential for fast Winograd and Fourier convolution on AVX-512 CPUs. See https://NN-512.com