| ▲ | measurablefunc 6 hours ago |
| Why can't it be done w/ AI? Why does it need to be maintained w/ manual programming? Take the ROCm specification, take your CUDA codebase, let one of the agentic AIs translate it all into ROCm or the AMD equivalent. |
|
| ▲ | jsheard 6 hours ago | parent | next [-] |
| The article is literally about how rote translation of CUDA code to AMD hardware will always give sub-par performance. Even if you wrangled an AI into doing the grunt work for you, porting heavily-NV-tuned code to not-NV-hardware would still be losing strategy. |
| |
| ▲ | measurablefunc 6 hours ago | parent [-] | | The point of AI is that it is not a rote translation & 1:1 mapping. | | |
| ▲ | jsheard 5 hours ago | parent [-] | | > Take the ROCm specification, take your CUDA codebase, let one of the agentic AIs translate it all into ROCm ...sounds like asking for a 1:1 mapping to me. If you meant asking the AI to transmute the code from NV-optimal to AMD-optimal as it goes along, you could certainly try doing that, but the idea is nothing more than AI fanfic until someone shows it actually working. | | |
| ▲ | measurablefunc 5 hours ago | parent [-] | | Now that I have clarified the point about AI optimizing the code from CUDA to fit AMD's runtime what is your contention about the possibility of such a translation? | | |
| ▲ | bigyabai 5 hours ago | parent [-] | | There is an old programmer's joke about writing abstractions and expecting zero-cost. | | |
| ▲ | measurablefunc 5 hours ago | parent [-] | | How does that apply in this case? The whole point is that the agentic AI/AGI skips all the abstractions & writes optimized low-level code for each GPU vendor from a high-level specification. There are no abstractions other than whatever specifications GPU vendors provide for their hardware which are fed into the agentic AI/AGI to do the necessary work of creating low-level & optimized code for specific tasks. | | |
|
|
|
|
|
|
| ▲ | colonCapitalDee 4 hours ago | parent | prev | next [-] |
| No. This is far beyond the capabilities of current AI, and will remain so for the foreseeable future. You could let your model of choice churn on this for months, and you will not get anywhere. It will be able to reach a somewhat working solution quickly, but it will soon reach a point where for every issue it fixes, it introduces one or more issues or regressions. LLMs are simply not capable of scaffolding complexity like a human, and lack the clarity and rigorousness of thought required to execute an *extremely* ambitious project like performant CUDA to ROCm translation. |
| |
| ▲ | measurablefunc 4 hours ago | parent [-] | | I don't have a model of choice. I'm just going by what I hear on twitter from Sam Altman & the people who work for him. | | |
|
|
| ▲ | cbarrick 6 hours ago | parent | prev | next [-] |
| Has this been done successfully at scale? There's a lot of handwaving in this "just use AI" approach. You have to figure out a way to guarantee correctness. |
| |
| ▲ | measurablefunc 6 hours ago | parent [-] | | There are tons of test suites so if the tests pass then that provides a reasonable guarantee of correctness. Although it would be nice if there was also proof of correctness for the compilation from CUDA to AMD. |
|
|
| ▲ | bee_rider 5 hours ago | parent | prev | next [-] |
| The AI is too busy making Ghibli profile pictures or whatever the thing is now. We asked it to make a plan for how to fix the situation, but it got stuck. “Ok, I’m helping the people build an AI to translate NVIDIA codes to AMD” “I don’t have enough resources” “Simple, I’ll just use AMD chips to run an AI code translator, they are under-utilized. I’ll make a step by step process to do so” “Step 1: get code kernels for the AMD chips” And so on. |
| |
| ▲ | measurablefunc 5 hours ago | parent [-] | | The real question is whether it will be as unprofitable to do this type of automated runtime translation from one GPU vendor to another as it is to generate Mario clips & Ghibli images. |
|
|
| ▲ | j16sdiz 5 hours ago | parent | prev | next [-] |
| The same as "Why just outsourcing it to <some country >" AI aint magic. You need more effort to manage, test and validate that. |
| |
| ▲ | measurablefunc 5 hours ago | parent [-] | | Isn't AGI around the corner? If it is then this is a very simple problem that should be solvable w/ existing pre-AGI capabilities. | | |
| ▲ | j16sdiz 5 hours ago | parent | next [-] | | I am not saying this is impossible, but I am down voting this because this is _not an interesting discussion_. The whole point of having an online discussion forum is to exchange and create new ideas. What you are advocating is essentially "maybe we can stop generating new ideas because we don't have to. we should just sit and wait"... Well, yes, no, maybe. but this is not what I expect to get from here. | | |
| ▲ | measurablefunc 5 hours ago | parent [-] | | You can do whatever you want & I didn't ask you to participate in my thread so unless you are going to address the actual points I'm making instead of telling me it is not interesting then we don't have anything to discuss further. |
| |
| ▲ | j16sdiz 5 hours ago | parent | prev | next [-] | | So, your strategy for solving this is: Convert it to another harder problem (AGI). Now it is somebody else (AI researcher)'s problem. This is outsourcing the task to AI researchers. | | |
| ▲ | measurablefunc 5 hours ago | parent [-] | | They keep promising that this kind of capability is right around the corner & they keep showing how awesome they are at passing math exams so why is this a more difficult problem than solving problems in abstract algebra & scheme theory on humanity's last exam or whatever is the latest & greatest benchmark for mathematical capabilities? |
| |
| ▲ | nutjob2 4 hours ago | parent | prev [-] | | > Isn't AGI around the corner? There isn't even a concrete definition of intelligence, let alone AGI, so no it's not. That's just mindless hype at this point. | | |
| ▲ | measurablefunc 4 hours ago | parent [-] | | Are you trying to say that Sam Altman & Elon Musk have been less than honest about their prognosis of luxurious & automated future utopias? |
|
|
|
|
| ▲ | Blackthorn 6 hours ago | parent | prev | next [-] |
| I don't know why you're being downvoted because even if you're Not Even Wrong, that's exactly the sort of thing that has been endlessly presented by people trying to sell AI as something that AI will absolutely do for us. |
| |
| ▲ | measurablefunc 6 hours ago | parent [-] | | Let's see who else manages to catch on to the real point I'm making. | | |
| ▲ | bigyabai 2 hours ago | parent [-] | | It's hard to catch-on to a deliberately dishonest pretense. You could clone 10,000 John Carmacks to do the job for you, Nvidia would still be a $5 trillion business next time you wake up. | | |
| ▲ | measurablefunc an hour ago | parent [-] | | There is nothing dishonest in what I wrote. If you want to really talk about dishonesty then the people you should be addressing are Sam Altman & Elon Musk. | | |
| ▲ | bigyabai 39 minutes ago | parent [-] | | I'm not talking to them. I am responding to you - your sardonic piss-take is against HN guidelines and written in bad-faith. | | |
| ▲ | measurablefunc 32 minutes ago | parent [-] | | I recommend you move on & refrain from responding to any of my posts in the future. If you're so concerned about the guidelines then reach out to the moderators & make your case, otherwise, I further recommend you save yourself the keystrokes. |
|
|
|
|
|
|
| ▲ | bigyabai 6 hours ago | parent | prev [-] |
| Because it doesn't work like that. TFA is an explanation of how GPU architecture dictates the featureset that is feasibly attainable at runtime. Throwing more software at the problem would not enable direct competition with CUDA. |
| |
| ▲ | measurablefunc 6 hours ago | parent [-] | | I am assuming that is all part of the specification that the agentic AI is working with & since AGI is right around the corner I think this is a simple enough problem that can be solved with AI. |
|
