NVIDIA needs to be broken up
You look at that and want to take a sledgehammer to a golden goose? I don't get these people
They saw there was nascent compute use of GPUs, using programmable shaders. They produced CUDA, made it accessible on every one of their GPUs (not just the high-markup professional products) and they put resources into it year after year after year.
Not just investing in the product, also the support tools (e.g. a full graphical profiler for your kernels) and training materials (e.g. providing free cloud GPU credits for Udacity courses) and libraries and open source contributions.
This is what it looks like when a company has a vision, plans beyond the next quarter, and makes long-term investments.
Big ships take time to course correct. Look at their hiring for AI related positions and release schedule for ROCm. As well as multiple companies like mine springing up to purchase MI300x and satisfy rental demand.
It is only May. We didn't even receive our AIA's until April. Another company just announced their MI300x hardware server offering today.
That is obvious and the signs are there showing that they are working on it. What more do you expect?
George Hotz tried to get a consumer card to work. He also refused my public invitations to have free time on my enterprise cards, calling me an AMD shill.
AMD listened and responded to him and gave him even the difficult things that he was demanding. He has the tools to make it work now and if he needs more, AMD already seems willing to give it. That is progress.
To simply throw out George as the be-all and end-all of a $245B company... frankly absurd.
Also, if the consumer GPUs are hopelessly broken but the enterprise GPUs are fine, that greatly limits the number of people that can contribute to making the AMD AI software ecosystem better. How much of the utility of the NVIDIA software ecosystem comes from gaming GPU owners tinkering in their free time? Or grad students doing small scale research?
I think these kinds of things are a big part of why NVIDIA's software is so much better than AMD right now.
I’d say it simply dials it down to zero. No one’s gonna buy an enterprise AMD card for playing with AI, so no one’s gonna contribute to that either. As a local AI enthusiast, this “but he used consumer card” complaint makes no sense to me.
My hypothesis is that the buying mentality stems from the inability to rent. Hence, me opening up a rental business.
Today, you can buy 7900's and they work with ROCm. As George pointed out, there are some low level issues with them, that AMD is working with him to resolve. That doesn't mean they absolutely don't work.
https://rocm.docs.amd.com/projects/install-on-linux/en/lates...
One way to improve the flywheel and make the ecosystem better, is to make their hardware available for rent. Something that previously was not available outside of hyperscalers and HPC.
I didn't do that, and I don't appreciate this misreading of my post. Please don't drag me into whatever drama is/was going on between you two.
The only point I was making was that George's experience with AMD products reflected poorly on AMD software engineering circa 2023. Whether George is ultimately successful in convincing AMD to publicly release what he needs is beside the point. Whether he is ultimately successful convincing their GPUs to perform his expectations is beside the point.
Except that isn't the point you said...
"there's no hope of them becoming serious contenders in AI without some major changes in AMD's priorities"
My point in showing you (not dragging you into) the drama, is to tell you that George is not a credible witness for your beliefs.
My point is as I wrote in both posts. George was able to demonstrate evidence of poor engineering which "reflected poorly on AMD". From this I could form my own conclusion that AMD aren't in an engineering position to become "serious contenders in AI".
The poor software engineering evident on consumer cards is an indictment of AMD engineers, and the theoretical possibility for their enterprise products to have well engineered firmware wouldn't alleviate this indictment. If anything it makes AMD look insidious or incompetent.
GPU compute is already broken up - there is a supply chain of other cooperating players that work together to deliver GPU compute to end users:
TSMC, SK hynix, Synopsys, cloud providers (Azure/Amazon etcetera), model providers (OpenAI/Anthropic etcetera).
Why single out NVidia in the chain? Plus the different critical parts of the chain are in different jurisdictions. Split up NVidia and somebody else will take over that spot in the ecosystem. This interview with Synopsys is rather enlightening: https://www.acquired.fm/episodes/the-software-behind-silicon...
How does the profit currently get split between the different links? Profit is the forcing variable for market cap and profit is the indicator of advantage. Break up NVidia and where does the profit move?