Intel CTO suggests using AI to port CUDA code to – surprise! – Intel chips

Saddled with a bunch of legacy code written for Nvidia's CUDA platform? Intel CTO Greg Lavender suggests building a large language model (LLM) to convert it to something that works on other AI accelerators – like maybe its own Gaudi2 or GPU Max hardware. "I'll just throw out a challenge to all the developers. Let's use LLMs …

[elsergiovolador]

Training

You most likely do need to port CUDA code to Intel chips first, so the AI can be trained on it and then port the CUDA code to Intel chips.

[_olli]

TF support, anyone?

Very very small fraction of people that develop AI actually develop CUDA code directly, instead they use libraries such as TensorFlow in between their frontend and the GPU

So how about getting that fine TensorFlow to EASILY work on those good intel GPU chips, and same to you AMD?

I have a suite of Amd, Intel and Nvidia GPUs in household and would be agnostic of which of them to use, yet Nvidia is the only brand that is easy to get working with common open-source AI frameworks. Amd has this rocm layer yet good luck getting it working unless using a very spesific OS distribution and/or kernel and/or GPU version.

[Peshman]

Re: TF support, anyone?

Isn't your problem with AMD a convergence issue? ATI used to be an indie GPU company. One merger later they're governed by their owners. NVidia are still steering their own ship and can be focussed.

[Yet Another Anonymous coward]

Re: TF support, anyone?

NVidia went all in on CUDA, the libraries and tooling are way beyond any of the alternatives and you can hire experts in CUDA.

Intel and AMD paid lip service to OpenCL support but only as a minimal box-ticking.

There are cross platform tools that will compile to CUDA or OpenCL but if you happen to need massively parallel GPU compute for anything requiring high performance you are going to end up tuning the resulting code, and only NVidia have decent tooling

[Snowy]

Re: TF support, anyone?

Paraphrasing what someone else said in a comment on a graphic driver thread, the hardware is just the start software is the other 95%.

[IgorS]

Niche player pushing niche solutions

If INTEL was serious about providing alternatives to NVIDIA, it would partner with the rest of not-NVIDIA GPU ecosystem.

In particular AMD, who has a very credible GPU solution on the market already!

And they have indeed invested a lot in software to support it, too.

But, no, INTEL has to go its own path and try to make its own walled garden.

(Only INTEL really supports SYCL right now!)

I get they want to make their own lock-in, but you cannot get that when you are a niche player.

They just can't get over the idea that they are not the dominant ones anymore/in this field.