Large corporate customers like Microsoft and Meta do not use CUDA. They all use custom software. AMD doesn’t have enough GPUs to sell them yet, that’s the real bottleneck.

That's a pretty big claim, that Microsoft and Meta have their own proprietary cuda-replacement stack. Do you have any evidence for that claim?

I'm guessing what they meant is that they use toolchains that are retargetable to other GPUs (and typically compile down to PTX (nVidia assembly language) on nVidia GPUs rather than go through CUDA source -- GCC and clang can both target PTX). For example XLA and most SYSCL toolchains support much more than nVidia.

Even then it's an insanely bold assumption that a company other than Nvidia could build a better framework than CUDA for compiling PTX. Especially since CUDA is already so much C-like. I've never seen anyone go deeper than that outside of academia.

How many customers/consumers will care about services are be built with CUDA?

If they need a ChatBot that uses a model with same accuracy and performance as on non-CUDA hardware, would they still want CUDA based hardware?

Who is going to build the architecture and compile the device specific kernels? You have to pay those people as well and you can save tons of money and time if you do it with cuda.

And this shouldn't be to hard if you know the ins and outs of the hardware and have a reasonable dev team. So why aren't they doing it?

> and have a reasonable dev team

probably this.

Microsoft recently announced that they run chatgpt 3.5 & 4 on mi300 on Azure and the price/performance is better.

https://www.amd.com/en/newsroom/press-releases/2024-5-21-amd...

I've used ChatGPT on Azure. It sucks on so many levels, everything about it was clearly enforced by some bean counters who see X dollars for Y flops with zero regard for developers. So choosing AMD here would be about par for the course. There is a reason why everyone at the top is racing to buy Nvidia cards and pay the premium.

"Everyone" at top is also developing their own chips for inference and providing APIs for customers to not worry about using CUDA.

It looks like the price to performance of inference tasks gives providers a big incentive to move away from Nvidia.

There are only like 3 AI building companies who have the tech capability and resources to afford that and 2 of them don't even offer their chips to others or have gone back to Nvidia. The rest is manufacturers desperately trying to get a piece of the pie.

Unless you develop in CUDA, you can easily train code (e.g. PyTorch) written for training on Nvidia hardware on AMD hardware. You can even keep the .cuda() calls.

In theory. But if you actually work with that in practice, you're already going to have a bad experience installing the drivers. And it's all downhill from there.