Broadcom Introduces 400G Ethernet Adapters for AI Data Centers

Broadcom has unveiled a new range of 400G Ethernet adapters designed to address network bottlenecks in AI data centers.

These highly scalable, low-power PCIe Gen 5.0 adapters are specifically engineered for the demands of AI processing, featuring advanced technologies such as a third-generation RoCE pipeline, low-latency congestion control, and telemetry capabilities. They are ideal for high-bandwidth, high-stress environments typical of AI infrastructure.

However, these are not to be confused with smartNICs. Jas Tremblay, Vice President and General Manager of Broadcom’s Data Center Solutions Group, clarified that their product differs from traditional smartNICs.

“These are high-performance NICs tailored for AI,” Tremblay explained. “Unlike smartNICs, they don’t have a large multi-core CPU. They are purely hardware offload solutions designed for optimal performance and low power usage in AI contexts. These processors are not meant for application offloading but focus on providing high connectivity, performance, and low latency with congestion control.”

Broadcom offers these processors both as standalone firmware on network cards and as chiplet options for chiplet-based designs, marking an industry first. This means that server processors using chiplet SoC designs could potentially integrate Broadcom’s networking chip into their systems. For instance, AMD’s Epyc processors use a chiplet design combining cores, memory, and I/O on a single die, and could theoretically incorporate Broadcom’s technology. While there is no indication that AMD is doing this, the option is available.

There is a distinction between the standalone chip and the chiplet version; the standalone chip boasts more features, including support for various virtual machines, while the chiplet is optimized for performance. This allows customers to choose based on their specific needs.

Additionally, Broadcom is launching Ethernet adapters tailored for copper wiring, capable of transmitting data over distances up to five meters. While fiber optics have their advantages, they are highly sensitive to heat, which is a significant issue in AI servers due to their high temperatures. Copper wiring, therefore, is a more practical choice for AI servers, and Broadcom has extended the distance 400G data can travel using copper.

The 400G PCIe Gen 5.0 Ethernet series (BCM57608) is widely available from several server vendors as well as directly from Broadcom.