It sort of does need an alternative, but also, on the AI front, I don't think Cuda is really that important anymore.
It's finally teetering & I think will really be tilting hard soon, but more generic systems are happening. OpenXLA has just taken over a huge part of the AI regime, and it's IREE representation has all kinds of different targets.
Cuda was more important when there weren't well established solid frameworks. For sure, there's plenty of places in computing where we do need general purpose low level GPGPU capabilities. And AMD and HIP leave a bit to be desired (the whole having to recompile the work again for each piece of hardware is a bitter pill especially). But in many places I think the hardware WI have very different means to be very very well targeted, rather than devs hand writing c++.
They’ve made a lot of promises over the years and never committed enough resources to deliver on most of them. They have made progress, no denying that, but it’s been slow. They have had to refocus several times over the years due to the moving hardware landscape and just how slow the project is going. There’s a history of failure to deliver on promises and so they need to start meeting expectations in order to build some trust in the project management.
It's finally teetering & I think will really be tilting hard soon, but more generic systems are happening. OpenXLA has just taken over a huge part of the AI regime, and it's IREE representation has all kinds of different targets.
Cuda was more important when there weren't well established solid frameworks. For sure, there's plenty of places in computing where we do need general purpose low level GPGPU capabilities. And AMD and HIP leave a bit to be desired (the whole having to recompile the work again for each piece of hardware is a bitter pill especially). But in many places I think the hardware WI have very different means to be very very well targeted, rather than devs hand writing c++.