NVIDIA Partners with TSMC on CPO Switches, Production Ramp in H2 2025

AI chip giant NVIDIA is not only conquering the computing market but also targeting data transmission as its next technological battleground. Gilad Shainer, Senior Vice President of Networking at NVIDIA, stated today that the company has collaborated with TSMC to develop the COUPE silicon photonics packaging platform and has begun shipping next-generation co-packaged optics (CPO) switches. Production capacity will be expanded in the second half of this year.

Speaking to the media at NVIDIA's enterprise product demonstration area, Shainer pointed out that as the computing scale of AI Factories grows explosively, the transmission bandwidth and power consumption between data center racks have become critical factors affecting AI performance.

To maximize the performance-per-watt ratio in AI factories, NVIDIA is fully advancing CPO technology. This breaks away from the traditional structure of placing optical transceivers outside the switch, directly integrating and packaging the optical engine with the switch chip to minimize power consumption and transmission distance.

TSMC plays a decisive role in the packaging process. Shainer revealed that NVIDIA is working closely with TSMC, utilizing TSMC's new COUPE (Compact Universal Photonic Engine) silicon photonics packaging platform. This innovation allows for the highly reliable and flexible packaging of optical engines, switch chips, and related components together.

"This allows us to move to mass production," Shainer emphasized. With the increasing prevalence of scale-out architectures, CPO technology will be seen everywhere in future AI data centers. The new Spectrum-X CPO switch, featuring the most advanced CPO technology and a throughput of up to 400 Tb/s, has already begun shipping to key partners, with production capacity expected to expand in the second half of this year.

Notably, amid the market's狂热 pursuit of "all-optical networks," Shainer also offered a pragmatic perspective. He said NVIDIA's networking strategy is very simple: "Use copper cables wherever possible."

Shainer explained that copper cables are cost-effective, extremely reliable, and have low power consumption. Therefore, for short-distance scale-up within a rack, such as connecting GPUs via NVLink technology, copper cables remain the preferred choice. However, when it comes to connecting hundreds or thousands of GPUs for long-distance scale-out between racks, that is where CPO and optical networks truly shine.