Source
- URL: https://x.com/cmuratori/status/2060486879263555907
- Author: Casey Muratori (@cmuratori)
- Posted: 2026-05-29 22:22:08
Branch
1/ @cmuratori
@Jonathan_Blow
What would that actually mean, in terms of hardware? I assume you don’t mean that the GPU would have a CPU on package, because that’s really no different than a Strix Halo or Core Ultra, etc.
2/ @Jonathan_Blow
@cmuratori
I mean the GPU is certainly a turing machine that could just run everything. But the amount of friction to overcome to make this happen is pretty close to infinity.
3/ @AgileJebrim
No actual physical hardware is a Turing machine…
Programming languages just simulate a TM as a VM target. Frankly it’s an absurd notion imo and it is a common misunderstanding that a TM is even required to be able to compute anything. TMs can’t even compute when they’ll halt when other types of machines can. 😂
4/ @Jonathan_Blow
@AgileJebrim @cmuratori
Sir, this is a Wendy’s.
5/ @cmuratori
@Jonathan_Blow
It’s a turing machine but it’s in-order, 1/3rd the clock rate, and has much higher L2 cache latency. So if you tried to execute any of the code people currently run on their computers, I imagine it would slow down by a massive factor? Not sure how to ballpark it - maybe 1/200?
6/ @Jonathan_Blow
@cmuratori
Yeah, I just think you could do some small things to reduce that factor by a lot, without having to make a whole CPU on the GPU. But it’s not what they are doing, so it doesn’t matter.
7/ @raggi
@Jonathan_Blow @cmuratori
Still not clear on the desire though? Do you want gpu style isa? You want uma with gpu memory? You want more parallelism closer to main?
8/ @Jonathan_Blow
@raggi @cmuratori
I want the massive massive shit pile to be slightly smaller.
9/ @raggi
@Jonathan_Blow @cmuratori
don’t we all :-)