Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Documentation for the AMD 7900XTX

github.com

Post reply on HN
Print view View original

101–110 of 154 posts

Re: Documentation for the AMD 7900XTX

#101

Post by viraptor »

lofaszvanitt wrote: post #92
ultra_nick wrote: post #31

Earlier quoted context omitted.

NVidia is making boatloads of money because their driver works and they have a software library called CUDA that accelerates neural networks. Nobody expects AMD to match them, but George thought they could at least write a GPU driver. If AMD can get that driver out, then George could provide a competitor to CUDA (for neutral networks only). They'd both make boatloads of money. However, AMD was less capable than expec…

Why is that, that AMD seemingly can't act like a sane individual would do?

There's a chance they (maybe specifically the lawyers) know something we don't. I mean, maybe they are absurdly incompetent in listening to feedback, while at the same time achieving technically great things in hardware. But after so many years and seeing all the AI money going to the competitor... That seems less and less likely every day.

Re: Documentation for the AMD 7900XTX

#102

Post by braiamp »

roenxi wrote: post #8

So for reference, this geohot is George Hotz who has a company Tiny Corp [0] in the space. Among a long list of things, he had a moment in the spotlight recently for a long rant [1] where he "gave up on AMD" because their drivers sucked more than he expected. The situation is reasonably complex - AMD have some programmers working on ROCm who seemed to be operating at standard fare when that is really not what AMD nee…

Common confusion. It isn't the kernel driver that has problems. Not their userland "drivers". It's their firmware inside the GPU's.

Re: Documentation for the AMD 7900XTX

#103

Post by braiamp »

MuffinFlavored wrote: post #33
georgehotz wrote: post #32

Earlier quoted context omitted.

I have had multiple replies from Lisa Su on both Twitter and by e-mail. It doesn't help. AMD is structurally incapable of fixing these issues. They don't even have a 7900XTX in CI. Crashing the firmware is so trivial there's no way anyone fuzzed anything. They debug by the application, adding mitigations at many layers of the stack to make it work. While this strategy is fine for the 20 mainstream games that come out…

Genuine question for you then, why are you investing your time reversing their firmware if it sounds like you think they're failing/will fail? Why not just "adopt" the golden child/golden standard NVIDIA now for your usecase, reverse that (if needed) and move on? Genuinely curious what you gain from trying to make AMD work/investing in AMD on your end here. Is it that you see a value proposition through the "muck" of…

> Why not just "adopt" the golden child/golden standard NVIDIA now for your usecase, reverse that (if needed) and move on?

Oh oh, I will answer you with the business side of the thing: because Nvidia sells exactly the same thing as Tinycorp. Also, Nvidia has shown that they will crush you if they get a whiff that you are doing something they don't like, for example, speaking with a competitor. With a business partner that gets you by the balls like that, it is not sound business strategy not to depend on them.

Re: Documentation for the AMD 7900XTX

#104

Post by orbital-decay »

viraptor wrote: post #101
lofaszvanitt wrote: post #92

Earlier quoted context omitted.

Why is that, that AMD seemingly can't act like a sane individual would do?

There's a chance they (maybe specifically the lawyers) know something we don't. I mean, maybe they are absurdly incompetent in listening to feedback, while at the same time achieving technically great things in hardware. But after so many years and seeing all the AI money going to the competitor... That seems less and less likely every day.

What might that be, for example?

"Organizationally unable to make competent software, perfectly able to make great hardware" seems to be the common case with hardware companies, if not de facto standard. Exceptions are rare.

Re: Documentation for the AMD 7900XTX

#105

Post by viraptor »

orbital-decay wrote: post #104
viraptor wrote: post #101

Earlier quoted context omitted.

There's a chance they (maybe specifically the lawyers) know something we don't. I mean, maybe they are absurdly incompetent in listening to feedback, while at the same time achieving technically great things in hardware. But after so many years and seeing all the AI money going to the competitor... That seems less and less likely every day.

What might that be, for example? "Organizationally unable to make competent software, perfectly able to make great hardware" seems to be the common case with hardware companies, if not de facto standard. Exceptions are rare.

Apart from "trying to implement this will cost us more in CUDA API copying lawsuits then it could earn", I don't know.

But it's not just that they can't make competent software. It's that everyone tells them they should try, that it looks like a pile of money ready to pick up, that people try doing it on their own... and AMD does nothing. They're not even taking the chance to fail/succeed. Can you imagine that Lisa Su doesn't get asked about this at least once a week?

Re: Documentation for the AMD 7900XTX

#106

Post by treffer »

One thing everyone should keep in mind about this NVIDIA/AMD battle right now: CUDA has been published for 16 years, it's been a huge push by NVIDIA to do GPGPU computation. I remember seeing it as a new thing in university back then, after the advanced shaders that were only available on NVIDIA.

NVIDIA pretty rightful has the lead there, because they worked and invested into it for something like 20 years (you could do pretty advanced shaders on NVIDIA pre-CUDA).

It only started to pay off recently, and especially with the AI hype (GPU mining was nice, too).

Now everybody is looking at the profits and goes like "OMG, I want a part of that cake!", either by competing (AMD / Intel) or by paying less for the cards (basically everyone else in the AI space).

But you have to catch up to 16 years of pretty solid software and ecosystem development. And that's only going to work if you have good enough hardware. NVIDIA did the hard work here. They have earned this lead.

I am saying this as someone who would rather not buy NVIDIA. I really wish I can soon throw 1-2 7900XTX into a machine and use it for LLMs without issues. But I would also bet that it takes at least a few more years to catch up, even with the massive global interest.

Re: Documentation for the AMD 7900XTX

#107

Post by llm_trw »

dannyw wrote: post #85
NicoJuicy wrote: post #43

Earlier quoted context omitted.

That's not really on how I see it. He's trying to do everything on AMD's consumer based GPU's, while he's comparing it with NVidia's enterprise GPU's with the same expectations. That's just not how it works in real life. But if he succeeds, yeah, he could save tons of money and his hardware for ML could actually be completed.

NVIDIA’s consumer GPUs are highly reliable and support the same CUDA capabilities as the enterprise products. This is something NVIDIA gets right. Meanwhile, ROCm isn’t even supported on certain lower end AMD consumer chips.

Rocm is only supported on the last generation of flagship graphics cards.

Which is a shame because the Radeon VII is still better than anything else when it comes to 64bit float computation.

Re: Documentation for the AMD 7900XTX

#108

Post by llm_trw »

earthling8118 wrote: post #68
glimshe wrote: post #44

I used to be an AMD diehard fan. Bought their processors and GPUs for close to 20 years until I simply couldn't take it anymore - 5 years ago I started buying NVIDIA GPUs with Intel CPUs and honestly never looked back. Sure I pay more, but it's worth paying more for reliability and software support. Anything else and I'm losing money out of ideology to give to a business that doesn't have its act together.

It's always intriguing to see other people's takes. I'm in nearly the complete opposite boat: in recent years I've switched to AMD and it feels like all of my hardware problems have gone away.

Heaven is using AMD cards for graphics and NVidia for computation.

Hell is the reverse.

I have an extra Radeon VII in my ML work station because I don't have to fight nvidia graphics drivers to get it to work. I have the NVidia cards so I don't have to fight AMD drivers to get ML drivers to work.

Re: Documentation for the AMD 7900XTX

#109

Post by CaptainOfCoit »

lofaszvanitt wrote: post #92
ultra_nick wrote: post #31

Earlier quoted context omitted.

NVidia is making boatloads of money because their driver works and they have a software library called CUDA that accelerates neural networks. Nobody expects AMD to match them, but George thought they could at least write a GPU driver. If AMD can get that driver out, then George could provide a competitor to CUDA (for neutral networks only). They'd both make boatloads of money. However, AMD was less capable than expec…

Why is that, that AMD seemingly can't act like a sane individual would do?

One individual doesn't suffer from the problem of being pulled in multiple different directions by multiple people. A company is not typically led by one person "dictator-style" but instead groups of people who try to make decisions together, sometimes not agreeing.

Re: Documentation for the AMD 7900XTX

#110

Post by devonkim »

fragmede wrote: post #73
jiggawatts wrote: post #72

Earlier quoted context omitted.

> don’t attempt to drown him in a bureaucratic swamp You weren't listening to the GP comment. He's expected to succeed while being drowned in the swamp . That's what most people mean and expect when they bring in an outsider. "Fix our problems, but without changing anything to do with management."

There's a difference between fixing technical problems and cultural problems. Hotz has proven to be a wizard at technical ones, while his skill with fixing cultural issues remains to be seen.

Any long-lasting solution requiring continuous investments is going to require more impact on the people side than the technical side. Hiring maybe one or two technical wizard asshats at a company may work for a bit but can’t scale in that eventually asshats will either drive away enough potential contributors or clash with another asshat. Any company actually needing said asshats to conduct business has a form of organizational tech debt in that they’ll need to spend a lot of resources managing around said persons in the end while said person’s ego continues to grow convinced they’re primarily responsible for successes.
Post reply on HN