3 ms·

I generally like Stratechery but this is not up to the usual standard. To the extent that CISC means anything it means “backwards compatible”. Modern ARM cores

I generally like Stratechery but this is not up to the usual standard. To the extent that CISC means anything it means “backwards compatible”. Modern ARM cores have multi-step decode and issue: to the extent that RISC means anything it means “we developed this with twenty year’s hindsight so we don’t have to do a lot in microcode yet.” In die area, in TDP, in wafer consumption, it’s just not a big cost. To the extent that Intel missed AI so did everyone: right now there’s one target and that’s NVIDIA, no amount of HBM/TC ratio tuning is going to change that. Like 3 of 17 Ray/Anyscale tutorials target Habana. The Gaudi stuff works fine if you’re willing to pay the “last year’s model” tax all non-CUDA does. And bigger picture, on semiconductors it’s the geopolitics people who are watching the bright line: semiconductors are a major power arena, the markets will get adjusted along those lines.

paulmd 2y ago
> 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 2y ago
  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".

paulmd 2y ago

> 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 2y ago
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".