Sounds interesting, I wonder if any of it will be available on consumer hardware. Making CUDA work across the stack seems to be the key thing that helped NVIDIA become the main purveyor of GPUs for AI. Apple seems to be doing well by the same approach.

Last time I looked, anything remotely useful here was only on the Xeon chips, which means there'll be very few open source and even fewer non-corporate folks doing anything interesting with it.

AMD seems to be better at that, maybe there's hope.