Google researchers Anna Goldie and Azalia Mirhoseini have developed an algorithm that can produce multiple new chip floorplans within a fraction of a second. This innovation is helpful to accelerate the advancement of artificial intelligence since the technology’s progress is interlinked with how fast and innovatively can the chips be designed. 

Computer chips are made through a process called chip placement, also known as chip floor planning, which requires 3-D design problem-solving. Traditionally, human engineers carefully design the configurations of hundreds to thousands of tiny components across for a layer, taking care that the minimum amount of wires are used to improve the chip’s efficiency. This design is then put through electronic design automation software for simulation and to verify effectiveness and performance. 

The verification takes at least 30 hours for a single floor plan. A complex machine learning algorithm has multiple layers of these intricate chip floor plans to be designed and verified.

However, due to the rapid evolvement of machine learning algorithms, the demand for new chip architecture has increased exponentially. While in the past, several algorithms have tried to optimise chip-floor planning, they’ve proved to be limited in their ability to produce good results in various important facets - such as chip’s power consumption, computational performance and coverage area. 

To combat this, Goldie and Mirhoseini trained their algorithm through reinforcement learning - a method of giving positive and negative feedback to teach complicated tasks to the algorithm. Whenever the algorithm would perform “well” it would be “rewarded”, and it would be “punished” for every “mistake”. 

The algorithm used the same function to evaluate the thousands of new chip floor designs it has created in a fraction of a second, and finally came up with a strategy for optimally placing the chip components. 

When verified, the AI algorithm not only made better floor plans as good to the ones designed by human engineers; they also taught the engineers some new tricks!