Somewhere in Google's latest tensor processing units sits a small recursion. Parts of each chip's floorplan, the arrangement of memory and logic blocks that determines how far every signal must travel, were produced by AlphaChip, a reinforcement-learning system that trains and runs on earlier TPUs. The chips help design their successors, which will help design theirs. It is the closest thing the hardware industry has to a compounding loop, and it comes with a healthy argument attached.
Your free preview ends here
Keep reading with a free account
Sign up and redeem two free articles per month. No credit card required.
Already a member? Log in
