Broadcom Claims First 5 nm Ethernet Adapter for AI Networking
To handle modern data loads, Broadcom has leveraged the 5-nm process node in its new Ethernet NIC to improve efficiency without sacrificing performance.
Broadcom has announced its latest high-performance Ethernet NIC solution purpose-built to increase data throughput and connectivity speeds of AI data centers.
The BCM57608 improves data throughput and efficiency, allowing data centers to enhance networking while minimizing rack usage. Image used courtesy of Broadcom
The BCM57608 is the “first Ethernet adapter built with 5-nm process technology.” Its 400 Gb/s aggregate data throughput, coupled with the improved thermal and power efficiency of the 5 nm process, enables greater rack density without requiring bespoke cooling solutions.
400G Built on 5 nm
Broadcom’s BCM57608 series of Ethernet controllers offers many specs that make it an attractive option for data center architects.
The PCIe-based NIC integrates seamlessly with existing PCIe switches for better integration with GPU-based servers. Image used courtesy of Broadcom
In addition to its speed, the BCM57608 series (datasheet linked) includes modern features that allow it to withstand data-center-grade network loads. Chief among these are remote DMA (RDMA) and RDMA-over-converged Ethernet (RoCE), which allow temporary spikes in data input to be mitigated without requiring software intervention. As a result, BCM57608-based Ethernet adapters offer greater performance in situations where congestion and contention are common.
The BCM57608 is available in PCIe- and OCP-NIC form factors with single- and dual-port options. These adapters are designed to be vendor-agnostic, allowing designers to use the NICs with a wide range of supporting hardware.
Modern Processes for Modern Solutions
Although modern processors (such as NPUs) and memory have evolved to better deal with the vast volume of AI data, designers still cannot reasonably store all training data on a single machine. Networking plays a critical role in modern high-performance computing, where many collocated computers work in tandem.
Developers must use expansive datasets to train next-generation models. This requires superior networking performance to prevent bottlenecks and continue growth. Image used courtesy of SimilarWeb
Networking devices like the BCM57608 must withstand network congestion to be reliable solutions for data centers. While data throughput is only one part of the equation for data centers and AI models, Broadcom feels confident that its new Ethernet NIC can have a measurable effect in improving data speeds and reducing bottlenecks.
