Home » NVIDIA Hopper has 13,000 instances of AI-designed circuits

NVIDIA Hopper has 13,000 instances of AI-designed circuits

We now know that the graphics chip NVIDIA Hopperswhich is the fastest artificial intelligence (AI) chip in the world, was created with the help of AI itself. The company itself revealed the information on its developer blog, where it reiterated the benefits and how it leveraged its AI capabilities to design its largest GPU yet, the NVIDIA H100.

As revealed, NVIDIA GPUs are mostly designed with the most advanced Electronic Design Automation (EDA) tools, but with the help of AI, which uses the methodology PrefixRLan optimization of parallel prefix circuits using deep reinforcement learning, the company can design chips smaller, faster and more efficient from an energy point of view, while offering superior performance.

Nvidia H100 - NVIDIA Hopper

“As Moore’s Law slows down, it becomes increasingly important to develop other techniques that improve the performance of a chip in the same manufacturing process. Our approach uses AI to design smaller, faster, and more efficient circuits that deliver more performance with each chip generation,” the company reveals on its official blog.

“Large arithmetic circuits have driven NVIDIA GPUs to achieve unprecedented acceleration for AI, high performance computing and computer graphics. Therefore, improving the design of these arithmetic circuits would be essential to improve the performance and efficiency of GPUs.

What if AI could learn to design these circuits? In ‘PrefixRL: Optimizing Parallel Prefix Circuits Using Deep Reinforcement Learning’, we show that AI not only you can learn how to design these circuits from scratchbut also circuits designed by AI they are smaller and faster than those designed by the most modern Electronic Design Automation (EDA) tools. The latest NVIDIA Hopper GPU architecture has nearly 13,000 instances of AI-designed circuitry.”

For reference, designing these nearly 13,000 AI-assisted circuits involves 25 percent reduction in area used regarding EDA tools. Of course, using PrefixRL is a very computationally heavy task, because for the physics simulation of each GPU you need 256 CPUs and over 32,000 GPU hours.

To remove this bottleneck, NVIDIA has developed Abductora proprietary distributed reinforcement learning platform that leverages NVIDIA hardware specifically for this type of industrial reinforcement learning.

About the author


Add Comment

Click here to post a comment

Your email address will not be published.