In late 2018, Google AI researchers Anna Goldie and Azalia Mirhoseini got the go-ahead to test an elegant idea. Google had invented powerful computer chips called tensor processing units, or TPUs, to run machine learning algorithms inside its data centers--but, the pair wondered, what if AI software could help improve that same AI hardware? The project, later codenamed Morpheus, won support from Google's AI boss Jeff Dean and attracted interest from the company's chipmaking team. It focused on a step in chip design when engineers must decide how to physically arrange blocks of circuits on a chunk of silicon, a complex, months-long puzzle that helps determine a chip's performance. In June 2021, Goldie and Mirhoseini were lead authors on a paper in the journal Nature that claimed a technique called reinforcement learning could perform that step better than Google's own engineers, and do it in just a few hours.

'We have already seen that there are algorithms or neural network architectures that don't perform as well on existing generations of accelerators, because the accelerators were designed like two years ago, and back then these neural nets didn't exist," says Google's Azalia Mirhoseini, "if we reduce the design cycle, we can bridge the gap." After studying chip designs, the neural net can produce a design for a Google Tensor Processing Unit in less than 24 hours that beats several weeks-worth of design effort by human experts in terms of power, performance, and area. Placement is so complex and time-consuming because it involves placing blocks of logic and memory or clusters of those blocks called macros in such a way that power and performance are maximized and the area of the chip is minimised. Heightening the challenge is the requirement that all this happen while at the same time obeying rules about the density of interconnects. Anna Goldie and Mirhoseini targeted chip placement it takes a human expert weeks of iteration to produce an acceptable design.