ST. LOUIS – AI-serving scale-up and scale-out architectures have become central to serving growing data center demands, with Broadcom one of many touting open interconnection models as being advantageous to framing those architectures.

Hemal Shah, a distinguished engineer in Broadcom’s Core Switching Group (CSG), told attendees during a panel discussion at this week’s SC25 event in St. Louis that open interconnect models like recently launched Ethernet for Scale-Up Networking (ESUN) for scale-up and Ultra Ethernet Consortium (UEC) specifications for scale-out.

Shah explained that these models allow for an open platform to decouple compute from networking, which is needed because of the obvious inability to have “one standard transport all the way to infrastructure.”

“But after talking through and doing all this, this is where ESUN was created,” Shah said of that scale-up model. “Let's separate out the underlying Ethernet infrastructure, make it transport independent, and then can be [Ultra Ethernet Transport] compliant interface or XPUs and want to have their own proprietary transport run over it, because that's probably the best optimized controller they can design for that XPU architecture. … That's why I think the whole separation of infrastructure from transport, that's how I see that down the road, they'll help with supporting both standard space as well as proprietary implementation on top of standard infrastructure.”

The Open Compute Project (OCP) unveiled ESUN last month, with a focus on designing large-scale data center infrastructures. This scale-up focus is targeted at connecting GPUs or accelerators within a single machine or tightly coupled cluster.

The initial working group includes participants from vendors like AMD, Broadcom, Cisco, Hewlett Packard Enterprise Networking (Juniper), and Nvidia.

ESUN’s focus on AI scale-up comes amid increased demand for AI services. Compared to InfiniBand-based interconnects, the decision to drill down on Ethernet for scale-up aims to use the open standard to help operators reduce costs and improve interoperability.

Scale and beyond

For scale-out, panel members touted UEC efforts. This is based on the recently unveiled UEC 1.0 specifications, which supports high-performance and scalable systems across all layers of the networking stack, including network interface cards (NICs), switches, optics, and cables.

A handful of vendors have already started to roll out UEC-compliant platforms, including Broadcom’s latest Thor Ultra NIC, AMD’s Helios, and Nokia’s 7220 Interconnect Router (IXR)-H6 series.

Combined, these efforts are targeted at providing an Ethernet-based networking option that can counter InfiniBand, which maintains a strategic hold on the market. This is due to that proprietary platform’s ability to provide “plug-and-play” networking that can meet high-performance computing (HPC) requirements.

Ethernet-backers counter that by noting the open interconnect standard is well-known to the vast majority of industry engineers and is vastly cheaper to interconnect hundreds of thousands of GPUs, which has helped drive its resurgence. A recent report from Dell’Oro Group suggests that increased interest in Ethernet-based offerings could drive nearly $80 billion in data center switch sales over the next five years.