Intel oneAPI 2023 Released - AMD & NVIDIA Plugins Available

[schmidtbag]

Interesting how they made plugins for AMD and Nvidia. It's a smart strategy:
Intel knows they have no chance of competing against CUDA if they make OneAPI proprietary or exclusive. Sure, they could support some other open standard, but if they can make their own that performs better than OpenCL, that would give them an edge. By giving AMD and Nvidia compatibility whether they asked for it or not, this allows more developers to consider adopting it. In another perspective, many devs avoid CUDA because it is vendor locked, or conversely, many people avoid software because it depends on CUDA. Meanwhile, so long as Intel optimizes their side of things, they will come out looking the best, which matters when there are alternatives to compare to. I'm sure AMD and Nvidia could make their own drivers/plugins if they really wanted to, but Nvidia has to much pride and AMD doesn't like spending that kind of R&D money, leaving Intel to look like the better deal overall.

[illuhad]

Sure, they could support some other open standard, but if they can make their own that performs better than OpenCL, that would give them an edge. By giving AMD and Nvidia compatibility whether they asked for it or not, this allows more developers to consider adopting it

That's true, but that is not an Intel idea originally. hipSYCL has been providing SYCL (which is the programming model for oneAPI) support for many devices including NVIDIA and AMD GPUs since 2018. And it can potentially even go through vendor compilers like AMD hipcc and NVIDIA nvc++ so that AMD and NVIDIA actually indirectly support SYCL, for the reason you mention.
In that sense, the AMD and NVIDIA support in the announcement is nothing inherently new - just support that is now available in another SYCL implementation.

[marlock]

Anything that runs nearly as well as native CUDA on Nvidia GPUs but has the advantage of also running over Intel and AMD has a fair shot at being widely adopted.

If it runs well over pure opensource drivers on AMD GPUs on Linux it's also in better standing than OpenCL itself.

And if it is as expressive as CUDA and less verbose and boilerplateful than OpenCL it's also in a great place.

Plus it's already a Khronos.org standard and it already works over multiple compatibility layers to ensure support without waiting for specific implementation! Pinch me, I must be dreaming?!

Oh god please make this the one people will pay attention to and adopt to finally end this decade of CUDA vs. OpenCL shenanigans where so many software "only run in hardware X" or "only run on driver Y that's worse than the out-of-the-box alternative at almost everything else" or "run better at hardware X because it's what devs spent more time optimizing than the cross-vendor alternative"