During the GTC conference, NVIDIA Chief Scientist and Senior Vice President of Research Bill Dally discussed how the company’s R&D teams are leveraging AI and machine learning to improve the design and efficiency of the company’s next-generation GPUs. Dally further discusses the use of machine learning and artificial intelligence to advance their goal of achieving better, more powerful GPUs.
NVIDIA discusses GPU design and the impact of AI and machine learning on future hardware
Dally gave an example of using AI and ML to augment inference to reduce a standard GPU design task from three hours to three seconds. These two methods optimize up to four slow and highly complex processes.
Dally drafted four important sections during the GTC session on GPU design and how AI and machine learning can make a big difference. Processes include controlling supply voltage drops, forecasting errors, etc., identifying and describing problems, and automating cell migration.
map voltage drop
This mapping of voltage drops allows NVIDIA to understand where power is going in next-generation GPU designs. Where once standard CAD tools could help in this process, the new AI tools used by NVIDIA can handle these tasks in seconds, which is a significant portion of that time. Implementing artificial intelligence and machine learning will increase accuracy by 94% and increase speed exponentially.
Parasitic prediction
Dally has a soft spot for using artificial intelligence to predict parasites. As a circuit designer, he spends hours with colleagues predicting parasitic effects during development. After NVIDIA completed the current test, they found that simulation errors dropped by less than 10%. This development improvement is excellent for circuit designers because it enables these developers to open up more creative and groundbreaking concepts in their designs.
Placement and routing challenges
Partitioning and routing challenges are important in the design of advanced chips because poor data flow can reduce efficiency exponentially. Dally said NVIDIA uses GNNs, or graph neural networks, to investigate and locate any problems and quickly find solutions that take up critical time during development.
Standard Cell Migration Automation
The migration of chips sometimes results in developers spending countless months developing without AI. Now, Dally says, “92 percent of the cell library can be done with this tool with no design rule or electrical rule errors” and that “in many cases, we end up with better designs.”
NVIDIA plans to prioritize artificial intelligence and machine learning across the company’s five labs. From Dally’s conference discussion, he hinted that we should see automatic standard cell migration included in its newer 7nm and 5nm designs, and that NVIDIA will include the Ada Lovelace series in these new projects.
