Google’s TPU is a custom ASIC specifically designed for machine learning with the accent on TensorFlow. It’s been in use now for a couple of years and has also been powering RankBrain, StreetView and Cloud Machine Learning.
However, a Californian based start-up called Cerebras has developed the ultimate ‘bad boy’ ASIC. Heralded as the largest computer chip in the world, it looks more like a tablet than a chip. Boasting 400,000 cores all linked by high speed connections, Cerebras maintains it will give less lag and lower power requirements than any other arrangement in machine learning.
Apparently a small quantity of these very large chips have already been shipped. One can only guess where?
Because of its significant size it is expected to see use in very specialised environments reducing the time taken to perform lengthy calculations from months to minutes.