> What baffles me is that Apple abandoned this completely sound theory for a risky (and entirely incorrect) bet on NPU hardware.
It wasn't a sound strategy. OpenCL was too little, too late, even though it was 15 years ago; CUDA was already dominant and OpenCL wasn't better in any way except being available on some hardware that wasn't as good as NVIDIA's, and OpenCL 2.0 a few years later was even more of a failure (NVIDIA basically refused to implement lots of new features and had enough leverage to force OpenCL 3.0 to make everything added after 1.2 optional). By contrast, Apple's NPU solved real problems for the iPhone in the domain of camera and computer vision features, operating within a reasonable power budget. Even today the NPU remains useful and superior to GPUs for some applications.
> they simplified their GPU hardware to specialize functionality better,
It wasn't a sound strategy. OpenCL was too little, too late, even though it was 15 years ago; CUDA was already dominant and OpenCL wasn't better in any way except being available on some hardware that wasn't as good as NVIDIA's, and OpenCL 2.0 a few years later was even more of a failure (NVIDIA basically refused to implement lots of new features and had enough leverage to force OpenCL 3.0 to make everything added after 1.2 optional). By contrast, Apple's NPU solved real problems for the iPhone in the domain of camera and computer vision features, operating within a reasonable power budget. Even today the NPU remains useful and superior to GPUs for some applications.
> they simplified their GPU hardware to specialize functionality better,
What features went missing, and when?