The issue is that hardware systems have evolved in a very dynamic way over the years. Initially GPUs were mostly fixed function hardware with proprietary APIs f

The issue is that hardware systems have evolved in a very dynamic way over the years. Initially GPUs were mostly fixed function hardware with proprietary APIs for each driver. Then we had Microsoft that was able to flex its platform power to build a vendor interface in Direct3D. Now we have Apple doing the same thing with Metal. The issue is going to be the ever increasing integration of heterogenous compute engines in unique ways. Google Tensor SoC, Apple A/M chips, etc. AMD is going to increasingly license their IP to build out custom SoC for companies like Samsung. The best chance to build a standard on top of all this change is to not view this as a "GPU abstraction" but rather a collection of compute engines. At this layer, the place where this happens will be compiler toolchain. IMO (as ex-NVIDIA compiler guy) the best effort happening right now is MLIR [1]. I don't know if dialec is the right abstraction but thats where things are at right now. The best example is is IREE [2] which is written at Google for mapping a lot of of the sophisticated translation inference on mobile for acceleration across GPU, neural engines, etc. [1] https://mlir.llvm.org https://mlir.llvm.org [2] https://google.github.io/iree/ https://google.github.io/iree/