If I’m reading this right, yes. Vulkan calls extension, extension translates the call and addresses the GPU directly in PTX.
Makes sense to me, CUDA itself isn’t extremely optimal. It was supposed to be the easy button to get fast parallel code executing on Nvidia GPUs but has mostly failed at both of those objectives.
For optimal you really have to get into the PTX language, which is perhaps a half step above Assembly.
But this really doesn’t help AMD, Intel etc’s position. They still have to do their own extensions and optimization. But would be nice for programmers to be able to consolidate on Vulkan for both ML/Gaming workloads.
I mean, it’s great, but why would nvidia ever want to give up their cuda monopoly?
It doesn’t. The extensions they’re using are Nvidia specific.
Does this mean they could feasibly move to vulkan-based ML with monopoly protected by their own extensions?
If I’m reading this right, yes. Vulkan calls extension, extension translates the call and addresses the GPU directly in PTX.
Makes sense to me, CUDA itself isn’t extremely optimal. It was supposed to be the easy button to get fast parallel code executing on Nvidia GPUs but has mostly failed at both of those objectives.
For optimal you really have to get into the PTX language, which is perhaps a half step above Assembly.
But this really doesn’t help AMD, Intel etc’s position. They still have to do their own extensions and optimization. But would be nice for programmers to be able to consolidate on Vulkan for both ML/Gaming workloads.