NVIDIA NVCell2 Powers 12X Faster Design

With advanced AI, NVIDIA is able to design cells for chips that power most electronics 12 times faster, according to a new preprint paper.

It’s a breakthrough that can be likened to playing Tetris at the microscopic level.
The paper, published on the Association for Computing Machinery’s website, details the cutting-edge method called NVCell2 that streamlines the design process of standard cells in Very Large Scale Integration, or VLSI chips.

VLSI chips integrate millions to billions of transistors and other electronic components onto a single microchip.

Standard cells are tiny and complex building blocks that are crucial for making devices like smartphones, computers and networking equipment faster and more efficient.

By intelligently arranging the miniscule building blocks, NVCell also boosts accuracy by 15 percent.

The method also improved performance of the designs by up to 7 percent, reduced energy usage by up to 8 percent, and decreased the space needed on chip by up to 4 percent when compared to traditional handcrafted designs. NVCell can handle more than a thousand chip parts at once.

The advancement could lead to more powerful devices advancing to production faster than ever.

The race to make everything smaller, including chip parts, presents incredible challenges to designers. The smaller the parts become, the harder it is to avoid mistakes.

The advancement is powered by novel transformer model-based clustering, which uses AI to intelligently group parts of a computer chip’s design to make the whole chip work better and easier to recreate.

It finds the best way to arrange the standard cells efficiently to allow for more components and reusable designs.

The paper was authored by NVIDIA’s Chia-Tung Ho, Ajay Chandna, David Guan, Alvin Ho, Minsoo Kim, Yaguang Li, and Haoxing Ren.