logoalt Hacker News

catapart • today at 5:17 PM • 3 replies • view on HN

Right. Sounds like that to me, too. But, again, I'm completely new to the topic, so what do I know?

If it is true that what is offering is some kind of "super-simd" processor that is built in a way that everything run through it must be similar to simd style code, but because it's hardware, it will run it much faster, then my questions are definitely targeted toward what the speed up is vs a highly-performant SIMD application. I would hope that their "coming next" stuff includes comparisons against the types of SIMD-boosted dense linear algebra you referenced. What is the speed up compared to that. And if it's less than (arbitrarily, say) 10x, how much can be gained by just having an LLM target breaking more stuff down in to multiplex-able instructions? But then, if it's more than that, my questions start turning towards "how much would it take to prototype" and "how much has to be changed from current tooling to produce the chips at scale" and "what is a good showcase target for an initial run? what would be something that could be made actually (in real dollars/hours) cheaper/faster/better by using this chip?"


Replies

rbanffy • today at 8:10 PM

From what I got, the magic is possible because the software is expressed as interaction networks and the CPU doesn’t need to resolve inter-instruction dependencies because the program makes them explicit.

fnimick • today at 7:40 PM

Isn't this just VLIW / EPIC that failed spectacularly in the early 2000s?

RicardoLuis0 • today at 5:31 PM

> If it is true that what is offering is some kind of "super-simd" processor

if it is, that has been tried before and failed (with itanium/etc), because the optimization burden on compilers means that they never end up getting the projected speeds over conventional execution