Compared to general GPUs, these chips are more specialized for DNN evaluations and may contain specialized components for matrix multiplications. One caveat mentioned in lecture chat is that these chips may have to be re-designed the if the algorithms for DNN change dramatically.

Specialized accelerator chips perform better than GPUs on neural network computations but the trade off is the hardware cannot be altered in the event that some hardware optimization updates need to be done (unlike FPGAs that are reprogrammable and can be patched with updates). Hence apple will keeps selling their latest chips.