Between the GPU, NPU and big.LITTLE cores, many phones have no fewer than 4 different power profiles they can run inference at. It's about as solved as it will get without an architectural overhaul.