Broadcom has introduced a new 400 Gigabit Ethernet Network Interface Card specifically  engineered for the demands of artificial intelligence (AI) and machine learning  (ML) workloads in data centers and cloud environments.

The BCM57608 400G NIC aims to meet AI's unconventional networking requirements like extremely high bandwidth utilization, intermittent data surges and long job durations. The new 400G NIC uses a 5nm process and fits in a PCIe Gen 5.0 slot on a server motherboard.The 400G NIC features 400 Gbps throughput, high-scale RDMA capabilities, and low power consumption optimized for AI workloads. It uses Broadcom's 112G SerDes technology to enable an extended 4-meter reach on passive copper cables using lower gauge 30AWG cabling. This extended reach reduces cabling costs.

Demand for 400G has been steadily rising since at least 2022 and with the growth of gen AI demand is accelerating further.

“This is the industry’s first 5nm 400G Ethernet NIC,” Jas Tremblay, vice president and general manager of the Data Center Solutions Group, Broadcom, told SDxCentral

Why AI needs 400G While other forms of data traffic including video have pushed the need for higher bandwidth, AI is pushing it further.

“AI network clusters require a significant amount of bandwidth and connectivity as LLM workloads need to be distributed across hundreds, thousands and eventually millions of nodes,” Tremblay said.

He added that the new 400G NICs will T be deployed in AI servers from major OEMs, as well as Hyperscale AI servers. Eventually he expects that  400G will be required for general purpose compute servers.

The 400G NIC is available in three forms - as a board product, chiplet, or IP - to fit different consumption models for integration flexibility.

How the Broadcom 400G embraces RoCE and RDMA A core element of what enables high-speed Ethernet at 400G are the remote direct memory access (RDMA) and RDMA over Converged Ethernet (RoCE) .

“RoCE/RDMA is a standards-based mode that is critical for AI traffic and workloads,” Tremblay said. “This is our third-generation product that supports this and it provides low-latency performance. ”

The new Broadcom 400G NIC enhances the RDMA protocol with new techniques optimized for AI that come from the Ultra Ethernet Consortium (UEC). Enhancements include optimizations such as out-of-order packet placement with in-order completion, selective ACK/retransmission, packet-level multipathing across multiple paths, and configuration-free congestion control algorithms.

By the power of Thor At the foundation of the 400G NIC is some new silicon from Broadcom.

Tremblay said that Broadcom is now launching its new Thor 2 ASIC, which is an SoC (system-on-a-chip (SOC)) that does all AI NIC processing. Broadcom is  also announcing a portfolio of board-level products based on this chip.

“It is a fully hardware accelerated design that optimizes high performance and low power,” he said.