Technology heavyweight Hewlett Packard Enterprise and optical IO startup Ayar Labs have embarked on a strategic collaboration to advance silicon photonics in the datacenter and integrate optical IO into a future generation of HPE’s Slingshot interconnect. Optical-based interconnects are essential to overcoming the bandwidth, latency, power and distance barriers of traditional electrical IO, the partners state, noting the need is especially pressing within the HPC and AI arenas.
HPE and Ayar Labs anticipate that they will be bringing products to market within a couple generations (they would not get more specific than that, but said to watch for additional disclosures over the course of the year). HPE, via its investment arm, Hewlett Packard Pathfinder, is backing the effort by investing in Ayar Labs. Paul Glaser, vice president and head of Hewlett Packard Pathfinder, called Ayar Labs’ optical IO technology “highly differentiated and critical to the evolution of high-performance computing architectures.”
The multi-year collaboration is focused on meeting future requirements for HPC and AI-oriented architectures with the goal of developing low-power optical-based interconnects to be implemented on the switch ASIC to start, and eventually (“likely”) the NIC. Enabling composable disaggregated architectures by also integrating optical connectivity between endpoint devices (CPUs, AI silicon, FPGAs) and pooled memory is another avenue the partners are exploring.
“Moore’s law was really successful in dumping more flops into a socket, and the challenge has always been: how do you move data in and out?” posed Hugo Saleh, senior vice president of commercial operations, managing director for Ayar Labs UK business, in an interview with HPCwire. “At Ayar Labs, that’s our value proposition; we [address] the performance, bandwidth, latency and power bottleneck that you get on moving data,” said Saleh.
The company has been developing and promoting in-package silicon photonics since its founding in 2015 and last year demonstrated the industry’s first Terabit per second Wavelength Division Multiplexing (WDM) optical link with its TeraPHY optical IO chiplet. Ayar has previously announced engagements with DARPA and Intel.
HPE’s Marten Terpstra referred to Ayar as a “clear leader” in the optical IO space, part of the larger field of silicon photonics, which uses light instead of electricity to transmit data. Terpstra, a senior director responsible for high-performance networking and silicon photonics connectivity solutions for HPE supercomputing systems, said it’s a matter of when, not if, electrical connectivity will become untenable.
“At some point, electrical effects are just not going to be sufficient anymore, cost-wise, power-wise, distance-wise, or latency-wise,” he said. “Generation over generation, we’re pushing the limits on how fast we can run the interconnect both between the switching infrastructure and the endpoints as well as between the switching infrastructures. And we know that somewhere in the next few generations, we are going to run into [the constraints of] simple physics.”
But it’s not just about the technological limits. “Every time we take a step up in signaling rates, the power and the cooling requirement and the complexity goes up, and the distance gets shorter,” said Terpstra. “So at some point, there is a physical limitation, but well before that, there’s a business limitation, because it’s just not cost-effective anymore.”
For all these reasons Terpstra said that for a coming generation of Slingshot as well as other interconnects, integrated optical IO becomes inevitable.
“At some point, it doesn’t become an option anymore,” said Terpstra. “Certain portions of connectivity don’t necessarily run into these physical limitations (as quickly), but others absolutely do and the only option that you have with those is to integrate optical IO.”
The crossing of future computing milestones could very well be tied to optical interconnects. The extreme-scale supercomputers brokered by the U.S. Department of Energy and other leadership centers around the world are targets for such technology. The three systems currently poised to be the United States’ first crop of exascale iron – Frontier, Aurora and El Capitan – are all based on the HPE Cray EX architecture, utilizing HPE’s current-generation Slingshot interconnect technology.
Ready for light speed
Reached for comment, Addison Snell, CEO of Intersect360 Research, offered his perspective on the advance of optical IO and how these vendors are positioning themselves.
“We have been hearing about optical interconnect in HPC for decades – I commented on them specifically in a post-SC08 HPCwire blog – but Ayar Labs appears ready to take them to the mainstream. The partnership with HPE is noteworthy on two fronts. It signals that HPE believes enough in Ayar’s technology future to want to claim it for Slingshot, and therefore, it creates additional funding that could help make it a reality,” said Snell.
“All our newest supercomputing technologies are ultimately limited by an old standard, the speed of light, and we’re not going to get a c 2.0 any time soon. Data movement is subject to immutable laws of physics. Optical interconnects have the potential to reduce latency in communication at every level of system architecture, from chip to cabinet.”
At this point, Snell said he does not foresee Slingshot becoming a standalone networking product outside HPE. “Rather, it is part of a trend,” he said, “along with Fujitsu’s Tofu and Atos’ BXI, back toward vendor-specific architectures. Managing components at scale is a critical differentiator for high-scalability segments like supercomputing and hyperscale.”