Broadcom has introduced a new series of 400G Ethernet adapters specifically tuned for resolving network bottlenecks when moving massive amounts of data around for AI processing. The new processors are equipped with a third-generation RoCE pipeline, low-latency congestion control technology, and telemetry functions. They are meant for a high-bandwidth, high-stress network environment associated with AI infrastructure.
But don’t call it a smartNIC. Jas Tremblay, vice president and general manager of the Data Center Solutions Group at Broadcom, said there is a difference between their product and what is generally accepted as a smartNIC.
“These are traditional performance NICs optimized for AI,” he said. “They don’t have a large, multi-core CPU running on them. They’re fully hardware offload, high performance, optimized for low power AI. This processor will not do application offload on it. It’s really [about] connectivity, high performance, low latency, congestion control.”
The processors are not only available as standard firmware installed on a network card, but also as a chiplet option for chiplet designs – which Tremblay said is an industry first.
What this means, in theory, is that a server processor that uses a chiplet SoC design could integrate the Broadcom networking chip into their SOC. For example, AMD uses a chiplet design with its Epyc server processors, combining chips that contain cores, memory, and I/O on a single silicon die. If AMD wanted to, it could incorporate Broadcom’s networking technology into the Epyc chip die. That’s not to say AMD is doing that – but it has the option to do so.
There is a difference between the standalone chip and the chiplet, however; the standalone chip is more feature-rich with support for many different types of virtual machines, for example, while the chiplet is more performance optimized. Customers can choose what suits their needs.
Separately, Broadcom is coming out with an array of Ethernet adapters specifically designed to handle copper wiring instead of fiber and transmit the data over five meters in length. For all of its positives, fiber-optic is extremely sensitive to heat, and AI servers are virtual ovens with their immensely hot GPUs running at full utilization. So copper is a better choice for an AI server, and Broadcom has increased the length 400G data can travel.
The 400G PCIe Gen 5.0 Ethernet family (BCM57608) is broadly available from multiple server vendors, as well as from Broadcom.
Source:: Network World