SANTA CLARA, California, September 24 (Reuters) — The powerful chips that run artificial intelligence applications are known for consuming significant amounts of electricity. On Wednesday, the world’s largest chip manufacturer revealed a new approach to make these processors more energy-efficient by using AI-driven software in their design process.

 

During a conference in Silicon Valley, Taiwan Semiconductor Manufacturing Company, which produces chips for Nvidia, demonstrated several methods aimed at increasing the energy efficiency of AI chips by up to ten times. For context, Nvidia’s current top AI servers can draw around 1,200 watts under heavy workloads, roughly equal to the continuous electricity usage of 1,000 homes in the United States.

 

TSMC plans to achieve these improvements through a new generation of chip designs that integrate multiple smaller chip components, called chiplets, which use different technologies into a single computing unit.

 

To fully utilize these advanced designs, chip designers are increasingly relying on AI-based software from companies like Cadence Design Systems and Synopsys. Both firms introduced new products on Wednesday, developed in close collaboration with TSMC. In some complex design tasks, the AI tools provided more effective solutions than human engineers and completed them much faster. Jim Chang, deputy director at TSMC’s 3DIC Methodology Group, explained that the AI tools could complete in five minutes work that would take a human designer two days.

 

Current chip manufacturing methods are approaching physical limitations, such as the challenges of transferring data through electrical connections. New approaches, including using optical connections to move data between chips, must become reliable enough for deployment in large-scale data centers. Kaushik Veeraraghavan, an engineer at Meta Platforms, emphasized in his keynote address that these challenges are not just engineering issues but fundamental physical constraints.