NVIDIA uses ML and AI technology to design GPUs

Chief scientist and senior vice president of research, Bill Dally led a team of about 300 people to share some NVIDIA R&D information at the previous GTC 2022, which involved developing, improving, and accelerating GPU designs using machine learning (ML) and artificial intelligence (AI) technologies.

Mapping the voltage drop will show designers what power supplies are used in new GPU designs, whereas traditional CAD tools take three hours to run and just three seconds using a trained AI model, which currently provides 94% accuracy, which is a greatly accelerated compromise; Bill Dally believes that AI tools can be very helpful in predicting parasitics, which can reduce the traditionally tedious and lengthy process, and the error is also low. Layout and wiring are very important for chip design. Once errors are made, the layout will be re-planned. AI technology is quite accurate in predicting this aspect. Even if it is not perfect, it can point out the problem area. Bill Dally said that switching chip design from 7nm to 5nm requires considerable effort, and 92% of cell libraries can be implemented with AI tools. The work that used to take 10 people nearly a year to complete can now be done in a few days by the GPU, which only needs to process the remaining 8%.

The demand side of NVIDIA research attempts to drive demand for NVIDIA products by developing software systems and technologies that run on GPUs. There are currently three different graphics research groups at NVIDIA to advance the development of computer graphics. There are also five different AI groups, because GPU application of AI technology is a big thing, and it is getting bigger. There are also groups working on robotics and self-driving cars, and there are multiple labs.