It's also complicated by the notion that raster performance doesn't directly translate to tensor performance. Apple and AMD both make excellent raster GPUs, but

It's also complicated by the notion that raster performance doesn't directly translate to tensor performance. Apple and AMD both make excellent raster GPUs, but still lose in efficiency to the CUDA's architecture in rendering and compute. I'd really like AMD and Apple to start from scratch with a compute-oriented GPU architecture, ideally standardized with Khronos. The NPU/tensor coprocessor architecture has already proven itself to be a bad idea.