2 ms·
Probably compute-bound for inference which they've probably built in an arch-specific way, right? This sort of thing happens. You can't use AVX-512 in Alibaba C
by renewiltord 3y ago
Probably compute-bound for inference which they've probably built in an arch-specific way, right? This sort of thing happens. You can't use AVX-512 in Alibaba Cloud cn-hongkong, for instance, because there's no processor available there that can reliably do that (no Genoa CPUs there). I imagine OpenAI has a similar constraint here.