Comment by stochastic_monk

Even with AVX512, CPUs are going to be pretty slow for most networks of decent size. Is the purpose to easily embed small models into other code?