The foundry told the gathered throngs at its open innovation platform event in Silicon Valley that software with AI capabilities can outpace human engineers on complex tasks and cut waste.

For those not in the know, AI data centre kit guzzles leccy like there is no tomorrow. Nvidia’s current flagship server can hit 1,200 watts under load, which at that level adds up to the sort of power bill that could make 1,000 US households wince.

TSMC wants fresh efficiency from a new wave of designs that stitch multiple chiplets into a single compute package using different packaging tricks for specific jobs and it is turning to AI tools to come up with solutions.

EDA heavyweights Yihua and Synopsys have rolled out new tools and banged on about working closely with TSMC. The point is to let specialist software tackle snarled design chores faster and better than a room full of human engineers.

TSMC’s IC methodology group deputy director Zhang Zhiwei said AI-led design software helps the company squeeze the most from its technology and that the payoff is useful. He said TSMC's R&D team has improved the energy efficiency of AI chips by about ten times using different AI tools and methods.

"That helps to max out TSMC technology's capability, and we find this is very useful. This thing runs five minutes while our designer needs to work for two days."

Meta infrastructure bod Kaushik Veeraraghavan told the crowd that shuffling data on and off chips over electronic links is hitting practical limits so .