Nvidia Deploys AI to Design Graphics Processors

Nvidia uses graphics processors and machine-learning algorithms to design new video accelerators. As reported by HPC Wire.

According to Nvidia’s Chief Scientist and Senior Vice President of Research, Bill Dally, AI can be effectively used in four key areas of GPU design:

• voltage-drop mapping;
• predicting parasitic effects;
• placement and routing issues;
• automation of standard-cell migration.

Voltage-drop mapping shows engineers how power is distributed in new processors. According to Dally, using the standard design method the necessary calculations take three hours. Applying the AI algorithm reduced this to three seconds with 94% accuracy.

“We can obtain very accurate voltage estimates much faster than with conventional tools, and in a very short time,” said Dally.

Graph neural networks were also used to analyze the placement and routing problem of processor components. According to Dally, improper execution of this condition will lead to data bottlenecks, akin to traffic jams on city roads, which would require a redesign of the chip layouts.

“It [the algorithm] shows problematic areas, and we can act accordingly and iterate very quickly, without the need for a full rerouting,” the scientist added.

AI-assisted automation of standard-cell migration can help accelerate adoption of new standards. Dally noted that moving from 7-nm to 5-nm process technology required substantial effort. Reinforcement learning helped speed up this stage and reduce the number of design-rule errors.

“This is a huge labour saving […]. In many cases we also achieved a better design,” said Dally.

In March 2022, Google introduced the PRIME algorithm, helping to design fast and compact processors for AI tasks.

In October 2021, the search giant spoke about using reinforcement learning to shorten chip creation times from several months to six hours.

In August, Samsung began using AI for automating the process of designing computer chips.

Subscribe to ForkLog news on Telegram: ForkLog AI — all the news from the world of AI!