Spacy training on Apple M1 vs. AMD Ryzen 5900X: 55% faster, 16x more efficient

that silly 5900x is drinking data through a straw. even if it had the raw oomph- it's more a touring car tham dragster- it'd still have 1/4th the memory bandwidth.

i am hella curious to know where the amx instructions run on the m1. with unified memory, there'd be seemingly little penalty to transparenty passing off to the gpu.