Carl Weiss

so we're talking billions of mac operations, that's a lot of compute hours - has anyone crunched the numbers on tpu vs gpu costs for large scale deployments, considering data egress fees too?

so they're importing pre-compiled cuda binaries to get around the proprietary compiler issue, that's clever but i have to wonder what the egress fees are going to look like when you're pulling down all those binaries from nvidia's servers