Neural Processing Units (NPUs)

Neural Processing Units are specialized processors designed to run neural network applications much faster than general purpose CPUs, or even GPUs. Google, with its Tensor Processing Unit (TPU), and Microsoft with its Azure NPUs, have been leading the way in this area.

Sophisticated AI models can now take huge amounts of compute power to train. It is rumoured that the latest (as of July 2020) GPT3 natural language model took close to $5M of compute time to train. Any technology that can make backpropagation (the algorithm used to train neural networks) run faster and more efficiently can save tons of time, electricity and money.

Here’s some further reading:

DNN Processor from Microsoft