The Massachusetts Institute of Technology (MIT) recently posted a blog highlighting how Google is using an artificial intelligence (AI) program to optimize the design of its AI chips. The Big Tech firm is using an innovative type of algorithm to improve its chip floor planning, the process of laying out components on a module.
So far, the corporation has had remarkable success with its atypical design methodology.
The Challenge of Contemporary Chip Architecture
Traditionally, chip floor planning involves an engineer laying out a new component configuration and verifying its design with software stimulations. Designers work to balance the need for efficiency, power consumption, and footprint with the latest component configurations. As such, the process, which includes 30 hours of automated testing, is very time intensive.
Manufacturers like Google have sought to optimize their component configuration tasks to meet the demand for new, highly sophisticated machine learning (ML) CPUs.
Indeed, firms are using powerful ML algorithms to optimize farming, medical treatment, solar utility development, clothing retail, and gamer humiliation. In addition, the global 5G rollout is expediting the automation of the manufacturing sector, a process that calls for new ML programs to manage autonomous operations. As different industries need programs capable of overseeing sector-specific tasks, there is a great need for customized tech stacks.
Two data scientists employed by Google recently took a novel approach chip planning; they used an AI to layout new AI chipsets.
Google’s AI Designs AI Chip Project
Google AI researchers Anna Goldie and Azalia Mirhoseini tasked a reinforcement-learning algorithm with designing new chip floor plans. Reinforcement-learning programs find new solutions to problems without a training data set. Instead, they function by addressing challenges in different ways via a reward/punishment mechanism until they find success.
Ultimately, Google’s experiment with AI chipset architecture provided a few different benefits.
Once active, Goldie and Mirhoseini’s program created thousands of new chip floor plans in a fraction of a second. The pair ran their AI’s layouts through an electronic design automation platform, which revealed its floor plans outperformed human-made configurations. The researchers also found the reinforcement-learning program discovered new methodologies engineers could use.
Given the outcome of Google’s experiment, it wouldn’t be a surprise to see other component makers employ ML algorithms to improve their chip architecture.