Nvidia starts shipping TSMC-backed CPO switches to tackle AI data centre bottlenecks

Nvidia Corp has announced that it has started shipping its next-generation co-packaged optics (CPO) switches, developed in collaboration with Taiwan Semiconductor Manufacturing Co. (TSMC), as the US tech giant seeks to address data transmission and power consumption challenges in AI data centres, according to a report by Focus Taiwan.

Speaking at Nvidia GTC Taipei on Wednesday (local time), Gilad Shainer, Senior Vice President of Networking at Nvidia, said the company has begun supplying its new Spectrum-X CPO switch to select partners.

According to the report, the switch incorporates Nvidia's most advanced CPO technology and delivers throughput of up to 400 terabits per second, he said, adding that production capacity is expected to increase during the second half of the year.


CPO switches are networking devices that integrate optical engines -- components that convert electrical signals into optical signals -- directly with switch chips, which route data between servers and processors, to transmit data more efficiently while reducing power consumption.

According to the report, Shainer said that the push comes as AI factories rapidly expand their computing capacity, making bandwidth and power consumption between server racks in data centers key factors affecting AI performance.

These server racks contain equipment such as servers, GPUs and switches that process and transmit data within AI data centres.

He noted that Nvidia is pushing ahead with CPO technology as a key approach to maximising the performance-to-power ratio in AI factories.

According to Shainer, CPO technology allows Nvidia to move beyond conventional designs that place optical transceivers outside switches by integrating and packaging optical engines directly with switch chips, thereby minimizing power consumption and transmission distances.

TSMC has played an important role in the packaging process, with the two companies working together to adopt the chipmaker's new COUPE (compact universal photonic engine, silicon photonics packaging platform, he said.

The innovation allows optical engines, switch chips and related components to be packaged together in a highly reliable and flexible way, paving the way for volume production, Shainer added.

At the event, Shar Narasimhan, Nvidia's director of product marketing for data center GPUs and AI training, said Taiwan's technology ecosystem has been a valuable partner in enabling technologies that support the company's expanding AI systems.

He said Taiwan's supply chain has two core strengths -- being "hard-working" and having "very advanced technology" -- making it an indispensable partner for Nvidia.

According to the report, Narasimhan said that with global demand for AI computing power already far exceeding available computing capacity, Nvidia needs Taiwan's ecosystem to help expand production for AI factories around the world.