Nvidia uses GPUs and machine learning algorithms to design new video accelerators. This writes HPC Wire.

According to Bill Dally, the company’s chief scientist and senior vice president of research, AI can be used effectively in four key areas of GPU design:

• voltage drop mapping;
• predict parasite phenomena;
• placement and orientation problems;
• automation of standard cell migration.

Voltage drop mapping shows engineers how power is distributed in new processors. According to Dalli, when using the standard design method, the necessary calculations are made in three hours. Using an artificial intelligence algorithm reduced this process to three seconds with 94% accuracy.

“We’re able to get very accurate stress estimates much faster and in a very short time than conventional instruments,” said Dalli.

Engineers also used graphical neural networks to analyze the problem of placing and routing processor components. Failure to do so would lead to “data jams” similar to congestion on metropolitan roads, and would require redesigning chip layouts, he said.

“HE IS [алгоритм] it shows problem areas and we can act on them and iterate over them very quickly without having to do a full redirect,” added the scientist.

Automating standard cell migration with AI can help accelerate adoption of new standards. Dalli noted that the transition from the 7nm to the 5nm chip manufacturing process required a lot of effort. Reinforcement learning helped speed up this step and reduce errors in design rules.

“This is a huge labor saver. […]. In most cases, we also have a better design,” Dalli said.

Recall that in March 2022, Google introduced the PRIME algorithm, which helps to develop fast and compact processors for processing AI tasks.

In October 2021, the search giant talked about using reinforcement learning to reduce the time it takes to build a chip from months to six hours.

In August, Samsung began using artificial intelligence to automate the process of developing computer chips.