I have been playing around in pytorch with an a770 16GB card and hit this error. The response seems to be https://github.com/intel/intel-extension-for-pytorch/issues/... that larger than 4gb allocations aren't supported even though the card is 16gb. I haven't seen a ton of stuff on intel arc for machine learning so wanted to share my experience
The AI boom has been going on for a couple years now, and Intel and AMD continue to be unwilling or unable to compete with nVidia head-to-head. Their level of incompetence borders on the implausible. Do they consider the AI developer GPU market not lucrative enough, or too far from their core competency (CPUs and cheap gaming cards only)?
Or... my pet conspiracy theory is that nVidia holds one or more critical patents, and Intel and AMD are quietly tiptoeing around it to prevent legal nuclear armageddon across the industry. This might explain why Intel and AMD both suddenly U-turned away from any kind of official CUDA compatibility: https://news.ycombinator.com/item?id=39344815
But I guess more probably, there are just too many pointy-haired MBAs in charge of the software, and thus billions get left on the table.
a) AI (formerly known as ML) was still a lucrative sector for hardware before LLMs were a big thing (nVidia has been well positioned for many years).
b) It will continue to be a big deal after the bubble pops. The Internet still ended up being a massively profitable sector even though the late-90's dot-com boom made it look stupid for a while; the same will happen with the AI boom.
c) Cryptocurrency was mostly useless, while AI/ML is very much not. ML has been important for a decade, it's just making more noise now.
