Google's Tensor Processing Unit
Google announced a chip of its own designed exclusively for neural networks, and said it had already been running in its data centres for a year.
Why it matters
Hardware was built for the method rather than the method fitted to existing hardware, and that changed the economics of large models.
The heart of the chip is a matrix multiply unit of 65,536 eight-bit elements; precision is deliberately reduced because networks tolerate it. Performance per watt exceeded contemporary processors and graphics accelerators by tens of times. The consequence runs wider than one company: specialised accelerators became an industry of their own, and access to them a condition of doing frontier research.