- TSMC unveils AI-driven chip design to boost energy efficiency.
- New chiplet-based architectures promise tenfold efficiency in AI computing.
- Cadence and Synopsys software outperforms engineers, accelerating complex chip design tasks.
At a Silicon Valley conference, Taiwan Semiconductor Manufacturing Co., the contract maker that produces chips for Nvidia, showcased various methods it aims to enhance the energy efficiency of AI computing chips by roughly 10 times. Nvidia's leading AI servers, for instance, can draw up to 1,200 watts while performing intensive operations, roughly equal to the energy consumption of 1,000 US households if operated nonstop.
The benefits TSMC aims to realize stem from a new generation of chip designs where multiple "chiplets” smaller components of complete computing chips utilizing various technologies are combined into a single computing package.
To utilize those technologies, the companies that create chips are increasingly depending on chip design software from vendors such as Cadence Design Systems and Synopsys, both of which introduced new products on Wednesday that were developed in close collaboration with TSMC.
Also Read: Japanese Delegation Eyes Semiconductor Hub in Dholera
In certain intricate chip design tasks, tools from TSMC's software partners identified superior solutions compared to TSMC's own engineers and accomplished this much more rapidly.
"That helps to max out TSMC technology's capability, and we find this is very useful," Jim Chang, deputy director at TSMC for its 3DIC Methodology Group, said during a presentation describing the findings. "This thing runs five minutes while our designer needs to work for two days."
The existing method of chip production is reaching barriers, including the capability to transfer data to and from chips via electrical links. Kaushik Veeraraghavan, an engineer from Meta Platforms' infrastructure team who delivered a keynote speech, stated that new technologies, such as transferring data between chips using optical links, must be made dependable enough for implementation in large data centers.