During the GTC conference, Bill Dally, NVIDIA’s chief scientist and senior vice president of research, discussed how the company’s research and development teams are utilizing AI and machine learning to increase the design and efficiency of the company’s next-gen GPUs. Dally further discussed the use of machine learning and artificial intelligence to advance their goals of a better and more powerful GPU.
NVIDIA discusses GPU design and the influence of AI and machine learning in tomorrow’s hardware
Dally gave an example of using AI and ML to increase inference to speed a standard GPU design task from three hours to three seconds. The two approaches have optimized up to four processes that ran slow and were highly intricate.
Dally drafted four substantial sections on GPU design and how AI and machine learning can significantly impact during the GTC conference. Processes include controlling drops in power voltage, anticipating errors and more, establishing and delineating issues, and cell migration automation.
Mapping Voltage Drops
This mapping drops in voltage allows NVIDIA to see where the power flow travels on the next-gen GPU designs. Where once standard CAD tools would help in the process, the new AI tools used by NVIDIA can process these tasks in seconds, which is a significant fraction of the time. Implementing AI and machine learning will increase accuracy by 94% and exponentially increase speed.
Dally has a soft spot for predicting parasitics utilizing artificial intelligence. As a circuit designer, he would spend long periods with his colleagues anticipating parasitics in the development process. With the current testing completed at NVIDIA, they saw a drop in simulation error, less than ten percent. This improvement in development is excellent for circuit designers, as it frees those developers to open up more inventive and breakthrough concepts in design.
Place and Routing Challenges
Zoning and routing challenges are significant to the design of advanced chips in that poor data flow can drop efficiency exponentially. Dally states that NVIDIA uses GNNs, or Graph Neural Networks, to investigate and locate any issues and quickly find solutions that would take crucial time out of the development process.
Standard Cell Migration Automation
The migration of a chip would sometimes cause developers to spend countless months in development without AI. Now, Dally states that “92% of the cell library was able to be done by this tool with no design rule or electrical rule errors” and that “in many cases, we wind up with a better design.”
NVIDIA plans to prioritize AI and machine learning in five of the company’s laboratories. From the conference discussions by Dally, he hints that we should see the inclusion of automated standard cell migration in their newer 7nm and 5nm designs and that NVIDIA will include the Ada Lovelace line in these new projects.