Crazy to not see a word about benchmarking/evaluation of the kernels.

IMO at this time when all performant GPU kernels are generated/hill-climbed, its the largest source of error we see.

Should be kernel run with a hot or cold cache in practice? Should your kernel be bitwise equivalent to a ref? What error margins should your kernel have to ref? How many times are you running the kernel to avoid thermal noise? Are you running on separate GPUs to avoid overfitting the current hardware? Can you cudagraph-compiling your kernel to remove launch overhead? Are you A/B B/A interleaving your kernel/reference runs to avoid running thermal bias?

As someone who loves being able to press a button to receive matmuls tuned to the 3090 by his feet, are there rules of thumb to work out how to set thresholds for evals?

Do I want to set a particular threshold on any individual op (or a layer for a given transformer)? e.g "for this test set, 99% of l2norms being within 1e-4 error of the golden reference kernel"

Or do I want to just shortcut all that and go with task based evals? eg "word error no worse than 1% greater than the reference speech-to-text model"?

I'm assuming the latter is the way to go in the first instance?

It depends on if you do training or inference and also at what scale.

For your task with the 3090, and I assume inference, I would like you suggest do end-to-end on target task.

Sometimes for example, a normalization layer can cancel out/fix an error from a previous layer, then it makes no sense to spend flops to save the numerics.

This is kind of a wild west in inference serving, you sometimes change exp to tanh, sometimes go full 4bit, sometimes remove full layers because you still perform fine on the end-task.

Numerics and correctness are more important for training kernels and backwards passes

Thanks! Interestingly I find that coding agents really don't now how to quantize a model to a given loose target. "Make it faster and accurate" is seemingly in distribution, "Make it _even_ faster and worse" is a good way to get Opus 5.5 to say "I tried nothing and I'm all out of ideas"

It’s not crazy at all, the blog post is clearly about human-guided optimization with LLMs rather than hillclimbing with an autonomous agent.

All of the above is as-relevant in human-guided optimization.

Yes, but they have nothing to do with GPUs and kernels. Those are general performance standards that would apply to any effort regardless of target.

Tell me more about cuda-graphs and async on CPU ;)

Sure, that might be worth discussing, I agree

Hmm, this is a good point. I tried to be fairly brief on this part and separated a lot of the profiling to a separate page since I usually use AI for benchmarking these days. I do think there are a lot of principles that humans should learn in practice, even if most of this is done with LLMs.

I'll update the post in the coming week with a section on benchmarking. The current post does have some details on measurement, but I agree that maybe it's not covered in the required depth.

It's a bit of a footgun to do benchmarking with AI models. If you worked with really performant GPU kernels (SOTA, better-than cuBLAS/FlashInfer level), you know that they reward hack:

Primer: https://ornith.ai/defense_kernel_hack.html

In-depth(follow links etc): https://x.com/myainotez/article/2037258418843771349

Parts relevant: https://www.coreauto.com/blog/when-ai-starts-writing-systems...

----

When you do this with benchmarking, the same thing happens but on a much deeper, shittier, harder to understand level. It will set up the caches to benefit your kernels, it will set up the timing to benefit your kernels, it will set up the evaluation order to benefit your kernels. I've seen half-cache flushes, which only benefit large shapes.

I wouldn't be surprised if you pointed a language model on your benchmarking with the comment above, and it found something that is reward-hacking your task.

Yes, they can reward hack. However, making sure the model knows what your requirements are (e.g., "do not quantize, make sure to evict, meet this set of specifications") can mitigate that, and simply looking at the code briefly (or having the LLM describe it back to you) can mitigate it quite easily. Often, such issues are a result of insufficient specificity by the user. If you tell it just to make the model faster, it'll do it and not assume additional requirements.

Same thing applies to benchmarking, a few guardrails goes a long way. As long as you aren't completely vague ("set up benchmarks", vs. "write tests over a variety of common shapes, be careful to control for environmental variables") the models do quite well. I've checked over the code it's written and it's all fairly reasonable. I do mention cache eviction in the article, this generally wasn't an issue and the model did this automatically on the first pass (back in Opus 4.6 days). I think you might not be giving them enough credit.

I'm not completely familiar with the writers of this article, but I've never had an LLM try to hack this, even sometimes telling me a target was impossible without quantization or approximation:

"Code based analysis already tells a detailed story, but there is an overlooked parameter: "boundaries of the hardware". " - the coreauto post

This project has spanned several models since Opus 4.5. I think even then such details were trivially known by them, thought they required much more handholding than they do now. I'm honestly not sure how the article writers ran into the hacks they did.

A good harness is great, but if it was that was all, everyone would have SOTA perf and we would be done with GPU kernels.

Look at this challenge:

https://openkernelchallenge.dev/#leaderboard

If you have the tokens, let your harness go at let's say the first bf16 problem, I am curious how far you will get with it (1.73x is the leader and 2.04x is roofline). I think if you manage to get above 1.63x, I would be really interested in learning how you did that.

Seems like the kernel portion of that competition is closed. I can probably compare whatever the serving winner is against JamInfer when it's out (seems like they start that track on the 12th). The top of the article shows the e2e times on some math RL tasks, with fp8 weights/activations/KV, comparing against SGLang on two models. They use the same concurrent batch size, quantization scheme, and the final lengths were about the same.

30B-A3B MoE

Jam - 67,241

SGLang - 44,696

+50.4%

8B Dense

Jam - 58,821

SGLang - 41,759

+40.9%

Math RL is pure decode, so kernels are the most major factor here.

Oh sorry I didn't know it closed. I understand, I was mostly curious because I hear this every now and then, people seemingly not running into the issues of reward-hacking with agentic harnesses which contradicts a lot of findings from people and myself in these competitions.

I don't think it's not a problem, and I'm sure it pops up extremely often in the case of competitions when things are won by margins of +0.01x and hundreds of people are chipping at it with agents for days. This is a very different scenario to what I assume most people might be doing, though.

Thanks for bringing up these points, it's definitely good food for thought to add to the blog!

The writing here is so clear. Do you use Llms to write?

My guess is that the LLMs helped with the diagrams but the writing is mostly or completely from a human. It’s quite good (and has a few typos :P)

I appreciate the compliment, yes this is spot on

The page design is the standard Claude design, what do you expect?

The post is actually mostly human written! I personally am not a fan of Claude's writing and don't even use Claude (my main subscription is Codex).

> On the H200s used for most of this work, the GPU's main memory is HBM, or high-bandwidth memory. Ironically, this is the slowest memory on the chip.

Nit: HBM is off chip, unlike shared memory and registers

Will update, thanks for pointing out.

Someone needs to write a book on this

I really liked AI Systems Performance Engineering by Chris Fregly

Programming Massively Parallel Processors is a classic (written by NVIDIA/CUDA engineers themselves), if you haven’t checked it out already