- The companies are developing the new module because their customers are asking for it
- Customers want to embrace more powerful technology for AI workloads with something they’re familiar with
- The new modules will increase bandwidth density from 200 Gbps-per-lane to 400 Gbps-per-lane.
OCP GLOBAL SUMMIT, SAN JOSE, CALIFORNIA — Cisco and Nvidia this week unveiled plans to develop an ultra-dense small form factor pluggable (USFP) optic module for AI data centers. The USFP is capable of 400 Gbps-per-lane and is expected to deliver 4x the bandwidth density and 1.3x the power efficiency of 200 Gbps-per-lane optics.
The companies are developing the USFP because their hyperscaler and neocloud customers are asking for it. The customers want to extend their use of octal small form factor (OSFP) modules but scale from 200 Gbps-per-lane to 400 Gbps-per-lane. With USFP, they can embrace next-generation optical technology for AI workloads without overhauling their data center infrastructure.
Fierce asked Will Eatherton, Cisco SVP of Engineering, to break down all this information for us.
He said pluggable modules are used in AI data centers to connect servers — which house the GPUs — to the switches. The modules “have the fiber optic pigtail coming out of them, and these plug into both servers as well as into the switches, and then the optic runs between, and that's how you connect up your cluster of hundreds of thousands to millions of GPUs,” he said.
The current OSFP technology is topping out, and data centers need an upgrade for the future.
In terms of the increase in “lanes,” this refers to the tiny copper pathways that carry electrical signals between the chip and the optical module.
Eatherton said others are working on different solutions to increase bandwidth for pluggable optic modules in data centers. For instance, Arista and Broadcom are working on a solution called XPO that involves liquid cooling going to the module. And co-packaged optics (CPO) is another alternative. But Eatherton said Cisco has gotten feedback from its customers that what they really want is something that's like the old thing, but in the next generation.
For their USFP, Cisco and Nvidia will work with a variety of vendors.
“These modules will have massive engineering,” said Eatherton. “You have an ecosystem. So, what we're announcing is we have simulated and tested and are ready for this module to move into the next stage. And Cisco and Nvidia will champion it.”
They will work with connector vendors and then go through a process of forming an open standard. And that’s why they announced the new product at the Open Compute Project (OCP) event, because OCP is all about open technologies that aren't proprietary or locked in.
Interestingly, even though Cisco and Nvidia are working together, and Arista and Broadcom are working on the competing standard XPO, Eatherton said that ultimately, everyone will probably support both standards just to cover their bases in terms of providing customers whichever product they want.
Although Cisco and Nvidia are not giving specifics about roadmaps or timelines for their USFP, Eatherton said everything is moving so fast it will take a matter of months, not years.
Its target market for OSFP are the AI hyperscalers, neoclouds, telcos around the world that are creating sovereign AI clouds, and even high end-enterprises.
RELATED STORIES:
Meta exec blows the OCP audiences’ mind with its data center innovations
Arista gets into scale-up game as accelerator landscape diversifies
Opinion: Can carriers finally move up the stack? AT&T, Cisco and NVIDIA say AI changes the equation