TAIPEI (Taiwan News) — TSMC is using AI-powered design software to target up to a tenfold improvement in the energy efficiency of AI computing chips, Reuters reported Thursday.

At a Silicon Valley conference on Wednesday, the world’s largest contract chipmaker outlined strategies to boost energy efficiency in AI computing chips. The company manufactures chips for clients including Nvidia.

Reuters noted that Nvidia’s current flagship AI servers can draw as much as about 1,200 watts during demanding tasks.

The approach relies on “chiplets” — smaller modules using different technologies — packaged together into a single computing unit, overcoming limits seen in conventional chips.

To implement these designs, chip developers are increasingly turning to AI design software from Cadence Design Systems and Synopsys. Both firms rolled out new AI-powered tools developed in close coordination with TSMC on Wednesday.

For some complex design tasks, the AI tools outperformed human engineers, delivering solutions in minutes that would otherwise take days. “This helps maximize TSMC’s technology potential,” said Jim Chang, deputy director of TSMC’s 3DIC Methodology Group.

Current chip manufacturing faces physical limits, such as moving data across chips using electrical interconnects. Optical interconnects may help, but they must be reliable for massive data centers, said Kaushik Veeraraghavan, an engineer at Meta Platforms.

“Really, this is not just an engineering problem,” Veeraraghavan said. “It’s a fundamental physics problem.”