At a conference in Silicon Valley, Taiwan Semiconductor Manufacturing Co (TSMC), the world's largest contract chip maker, announced a new strategy to reduce the power consumption of artificial intelligence (AI) chips.
The highlight of the plan is the collaboration with chip design software companies to apply artificial intelligence right in the process of creating chips.
According to TSMC, current AI chips, especially in Nvidia's high-end servers, can consume up to 1,200W when handling heavy tasks.
This is equivalent to the amount of electricity used by 1,000 US households in the same period, showing the huge pressure on energy that AI infrastructure is putting on. To solve the problem, TSMC aims to increase chip energy efficiency by 10 times thanks to the new generation of designs.
One of the solutions is to develop chiplets, which are small chips made using different technologies and combined in a complete computing package.
However, this process is extremely complicated and requires special support tools. This is the time for AI-based chip design software from Cadence systems Design and Synopsys to take the lead.
These two companies have just introduced new products, developed in close coordination with TSMC, to automate many previously capital-based design steps that had to be taken care of by engineers.
According to Jim Chang, Deputy Director of 3DIC Method Group at TSMC, in some tasks, AI tools have found more efficient solutions than human engineers and completed them in just a few minutes, instead of taking up to two days.
That helps us maximize technological capabilities while saving time and costs, Mr. Chang emphasized.
However, experts also warn that the chip industry is gradually reaching physical limits, such as the speed of data movement inside and outside the chip through power connections.
Kaushik Veeraraghavan, an engineer at Meta Platforms, said that the future will need new technologies such as optical data transmission.
However, to be widely applied, these solutions must be absolutely reliable in huge data centers.
By incorporating AI into the design stage, TSMC and its partners hope to open a more energy-efficient chip generation, reduce the burden of operating costs and contribute to the sustainable development of global AI infrastructure.
