The likes of Nvidia, AMD, and Broadcom have joined forces to form an industry consortium aimed at aligning open specifications for optical interconnects for scale-up.

The Optical Compute Interconnect’s Multi-Source Agreement group (OCI-MSA) brings together some of the biggest names building optical solutions, along with those set to use them in mammoth clusters.

Companies including Microsoft, Meta, and OpenAI have all committed to working with optics makers to help create an interoperable optical interface protocol to establish a “plug-and-play ecosystem,” allowing cluster operators to more easily disaggregate XPUs and scale-up switches via a common optical physical layer.

“The appetite for technology to address the power and cost constraints impacting AI cluster design is real and imminent,” Dan Rabinovitsj, Meta’s VP of hardware systems, explained. “We encourage adoption of this OCI protocol to decouple the need for larger scale-up domains from the limitations of electrical backplanes in high-performance AI clusters.”

The roadmap for the prospective specification will see OCI-MSA apply its focus to standardizing high-density interfaces to push speeds of up to 800 Gbps (800G) per fiber.

It’ll also be looking at scaling wavelength counts and data rates in a bid to reach 3.2 Tbps (3.2T) and beyond. That plateau has long been touted as the next-generational leap in interconnectivity, though optics vendors are only starting to ship 1.6T products this year. Just this week, Broadcom debuted a new 1.6T-focused digital signal processor – a chip that powers pluggable modules – which it claims will provide the underlying foundation for the eventual leap to 3.2T.

OCI-MSA wants to ensure that optics leveraging its standard include support for “massive scalability,” with its eventual spec aiming to support scale-up domains with even higher GPU counts and higher bandwidth per GPU. The group’s efforts include planned support for pluggables, on-board, and co-packaged optics (CPO).

The optics market is becoming increasingly fierce, with vendors vying to provide the interconnect technologies capable of boosting bandwidth for ever-increasing hyperscale-level AI clusters.

The formation of OCI-MSA marks a rare instance of vendors coming together – more so with Nvidia pledging to come aboard.

While the chip giant has historically focused its efforts on proprietary endeavors, it has become increasingly more open for some projects and products in recent years, notably lending its expertise to open-weight AI models; widening the circle of partners able to link with its hardware in rack-scale solutions via NVLink Fusion; and offering Ethernet-focused solutions alongside its long-touted InfiniBand offerings to a point where sales from each camp are “roughly on par.”

“Nvidia is a founding member of the OCI-MSA to establish a common optical standard across global AI infrastructures,” Nvidia networking SVP Gilad Shainer added. “By equipping best-in-class compute with state-of-the-art optics, the OCI-MSA can deliver the scale and performance required by the next era of super-intelligence.”

More industry collaboration; more scale-up

The group’s focus – sometimes referred to as vertical scaling, where capacity inside the server or rack is boosted by tightly coupling hardware – isn’t the only industry effort looking at scale-up architectures.

The Open Compute Project (OCP)’s Ethernet for Scale-Up Networking (ESUN) project is working to create a similar specification.

ESUN version 1.0, released just this week, was spearheaded by engineers from Microsoft and Meta. It looks to standardize Ethernet-based interconnects for large GPU domains with support for link-level congestion control mechanisms and improved link resiliency to reduce packet drops

Other scale-up networking-related efforts include the SUE framework, of which Broadcom was a leading voice in developing, which is aimed at boosting communication between XPUs in a multi-XPU system.

And for scale-out, the Ultra Ethernet Consortium (UEC) looks to support low-latency transport for high-throughput for cluster-to-cluster interconnectivity.