Comment by rahen
13 hours ago
Strictly speaking, this is very domain-specific and doesn't enable any performance that Triton couldn't already achieve (eliminating global memory round-trips via epilogue fusion is nothing new). The real takeaway is the design shift for LLM-driven codegen rather than handcrafted kernels.
LLMs are still bad at low-level hardware optimizations, but really good at high-level composition. Designing compiler abstractions with a restricted, composable API so an LLM can easily glue expert-written blocks together is a smart move. I suspect this will eventually become the norm for codegens as we move to agentic development.
>LLMs are still bad at low-level hardware optimizations, but really good at high-level composition.
I disagree. While yes they don't have all the architectural quirks of every GPU memorized, they are able to extract such optimizations from ISA docs and online guides. Now with 1M context available on frontier models, they can even fit the whole ISA definition in context (RDNA 3.5 here specifically) and spit out swathes of optimizations to try. The rest is just bruteforcing a single goal which they are extremely good at.
Or that's how simple it'll look until you have subtle bugs to solve somewhere deep in your stack.
Anyways, low-level hardware optimized GPU kernels has been an exceptionally good use case for agents in my opinion. They have far more trouble in other domains like doing GUI.
If you look at Anthropic's recent kernel optimization challenge, and the human leaderboard, humans are soundly beating Claude's best attempt.
I think the reason, as parent suggested, is that LLMs are great at composition (mash-ups/regeneration - this is essentially what they are trained to do), and not so great at innovation. How well they can do relative to a human, on a low level optimization problem, is going to depend on degree of similarity of the problem to things they were trained on and/or have access to.
The lack of fast GPU kernels written by AI does not lend credence to your theory.
Perhaps you missed work like https://crfm.stanford.edu/2025/05/28/fast-kernels.html ?
1 reply →
> and spit out swathes of optimizations to try.
Without any guarantees of functional correctness.
I imagine this is what’s already done for AI laying out hardware design.