"realistic" in this case means accepting that CUDA will never be replaced. Parts of it will get replaced (the easy bits, like matrix multiplication), but there

"realistic" in this case means accepting that CUDA will never be replaced. Parts of it will get replaced (the easy bits, like matrix multiplication), but there are thousands of esoteric CUDA features that don't make financial sense to rewrite. They exist for automotive, robotics/automation, industry, gaming and datacenter customers. Only the AI parts will get rewritten, and even that has taken Apple half a decade. I hope for alternatives and competition, but that's why I grieve OpenCL. Open CUDA alternatives are the only way you can drive Nvidia's customers away from them; Apple abandoned that dream a while ago. If they were still cooperating with Khronos and the rest of the industry, then there would be hope for dethroning Nvidia.