Apple has a dedicated "neural engine" which is designed as an inference NPU. Where as Google's TPU has a dual focus, both inference and training, which is a more complex design.