@PatrickToulme false
jalapeño kernels are in gluon -- before models helped us write them, we wrote them by hand no problem, machine was originally designed for human programming
not very different from pallas
as the hotchips talks mentions stock sol/astra is great at programming it
more on GPT/codex usage in Jalapeño and some of “where we’re going from there”…
as article says the models really grok’d XLS, but it’s made huge strides in systemverilog as lingua franca, if you last tried before Astra it’s time for another look…
There is a lot of deserved hype for how great OpenAI's Jalapeno ASIC and system will be. However, there is one major flaw that no one is talking about at all... (1/6)🧵
@siliconcodesign honestly hard to comment on others' adoption patterns
one thing i cited in the hotchips talk was that the hw team at openai has a unique mix of (a) believing things are possible (b) relentless focus on execution & how things'll actually get done
combo can be surprisingly rare!
@plugyawn Nice! Related q I've heard people ask is "when do we start sending programs to one another in english", where the nondeterminism of "what it hardens itself into" is a potential strength, as users can guide context & prefs. Challenges too in the correctness of the hardening, ofc.
just mean to say "stuff is trending this way, here's an example" than "the world is entirely changed overnight"
of course people can also just ask LLMs to yeet stuff at the source code level, it's the common case for what people do, just less rigorous/controlled
there have definitely been folks focused on eating the heuristics because they are well contained e.g. mlgo https://t.co/7CP2sPWLYK
i think you also see the surge in languages with strong formalisms leaning (🤨) on this kind of thing
@debadeepta to be clear that’s just an example of a functional description so i could relate it conceptually to an existing compiler pipeline people could be familiar with, we actually start with unoptimized gluon in our case (implicit in that performance climbing slide)
just because i'm seeing ad hominem attacks for some godforsaken reason
i was the one who was scrolling and @JordanNanos is 100% right
ai is what optimized it into its roofline achieving form
SemiAnalysis' Jordan Nanos reveals the OpenAI engineers scrolling their own kernel code had no idea what it did line by line, and it didn't matter because the AI does
"It's not slop, because it performs, but it's literally just AI-generated assembly basically puked out in Gluon, this low-level kernel programming language that they built on top of Triton, which uses this really interesting programming model."
"The point is, the guys who were scrolling through this code with us, it was kind of clear that they know a whole bunch about hardware, they know a whole bunch about the concepts in the system, and they just have no idea what this MLA kernel that they're showing us for DeepSeek actually does."
"You can go line by line and it's like nope, nope, nope."
"But it doesn't matter. The AI understands it, the AI tests it, and you see the results. It produces correct kernels that perform really well."
"And I think this is just a sign of what's to come again."
"That actual code is not necessarily something that a human has to reason about deeply, if the AI knows how to manipulate the data movement and the processing elements on the hardware that you've given it."
________
More takeaways from OpenAI's Jalapeno development: https://t.co/BXAzHU97fP
@moustafafayez@DilreetR@JordanNanos +1, we also have pipelines where we make it clean up kernels to make them nicer and more human readable/understandable, just weren't used in this particular case
@DilreetR@JordanNanos yeah, to be clear this is not at all my blanket endorsement to outsource understanding; as ever, time and attention are the precious resources, and the taste to choose what to onload and continue to ensure human intuition is involved in is the game we are all figuring out
@r32a_@positiveblue2@firesidealpha note it's all on the accelerator device, at which point you can really just do math, and any unnecessary math is going to be slowing you down so will not be selected -- it's not writing arbitrary host code or anything, just turning a functional description into kernels