[paulmd]

> In die area, in TDP, in wafer consumption, it’s just not a big cost.

And yet we live in a world where an AI 370 processor has less than 1/4 of the single-threaded perf/w of a “totally equivalent” M3 Max processor. One of those things can’t be true.

Who are you going to believe, Jim Keller or your lying eyes?

And it only gets worse when you realize that he didn’t vaguely say the thing people claim he said… people have spent 4 years now arguing against observable reality using a quotation from Keller that he didn’t even make lol.

[benreesman]

I find it amusing that it was the off-the-scales "SoC" people at Apple who pioneered the new best paradigm for vector compute. Not surprising mind you, the PA Semi people are terrifying, Apple is one of the best microarchitecture firms in history.

I find this amusing because we still call them "GPU"s. NVIDIA's revenue is like 85-90% datacenter IIRC, gaming is like, anti-trust insurance or something.

The "low end" just almost always eats the "high end".