HPC systems leader Hewlett Packard Enterprise and startup Ayar Labs, maker of chip-to-chip optical I/O connectivity, today announced a strategic collaboration to integrate silicon photonics within HPE’s high performance Slingshot fabric. HPE also placed a financial stake in Ayar’s development of optical I/O technologies.
“The collaboration will focus on Ayar Labs’ development of high speed, high density, low power optical-based interconnects to target future generations of HPE Slingshot, the industry’s only high performance Ethernet fabric specifically designed for HPC and AI solutions,” the companies said in their announcement. “By combining these technologies, the teams are well-positioned to design next-generation high-performance networking solutions and novel disaggregated system architectures that are critical for increased flexibility, efficiency, performance, and throughput to support data-intensive demands of future workloads.”
In a pre-announcement briefing yesterday, HPE declined to specify when Slingshot offerings incorporating optical I/O will come to market, indicating that it’s work with Ayar Labs is a long term project that “will take a few years.”
Traditional electrical-based high performance fabric technology, particularly fabrics scaled massively across high-end supercomputing clusters, poses increasing challenges to systems integrators like HPE as the HPC industry approaches the exascale era. As HPE said in today’s announcement, “electrical-based networking offerings will eventually reach bandwidth limits, creating challenges in latency and overall application performance.”
HPE-Cray’s Slingshot fabric may be running into some related challenges now. Slingshot, instrumental to Frontier, the first U.S. exascale system being installed at Oak Ridge National Laboratory, has reportedly run into performance problems that HPE and Oak Ridge engineers are working to overcome. Asked this week to provide details on Frontier’s Slingshot issues, HPE told us, “Regarding Frontier, we are not sharing updates at the moment but we will keep you abreast when we do.” While rumors persist of Slingshot-Frontiers problems, Oak Ridge spokespeople tell us they expect Frontier to be stood up, tuned and operational on schedule this year.
Longer term, HPE envisions future generations of HPC systems interconnects significantly enhanced by optical I/O, which is a silicon photonics-based technology that uses light instead of electricity to transmit data. The technology addresses both the need for higher data rates and improved energy efficiency (see “Composable HPC-AI at Scale: The Emergence of Optical I/O Chiplets“).
“Optical I/O uniquely changes the performance and power trajectories of system designs by enabling compute, memory and networking ASICs to communicate with dramatically increased bandwidth, at a lower latency, and at a fraction of the power of existing electrical I/O solutions,” said HPE. “The technology is also foundational to enabling emerging heterogeneous compute systems, disaggregated, pooled designs, and unified memory architectures that are critical to accelerating future innovation.”
Founded in 2015, Ayar Labs is in part based on 10-years of research and collaboration between MIT, UC Berkeley, and CU Boulder, and funding from DARPA via the POEM (Photonically Optimized Embedded Microprocessors) program. Intel has also invested in the company via Intel Capital.
Ayar Labs’s optical I/O chiplet, called TeraPHY, uses standard silicon fabrication techniques along with disaggregated multi-wavelength lasers to achieve high-speed, high-density chip-to-chip communication, with power consumption of between 1 and 10 picojoules. This contrasts with next-gen Ethernet solutions that are expected to be in the 20 to 30 picojoule range, according to Ayar Labs. The result is a platform incorporating a compute ASIC (be it CPU, GPU FPGA or some kind of SOC) and optical I/O chiplets integrated in a 2.5-D package.
“HPE is the leader in advanced systems architectures for HPC, AI, and networking,” said Charles Wuischpard, CEO of Ayar Labs and formerly of Intel. “This collaboration will accelerate the introduction of a whole new class of system architectures that overcome the existing limitations of traditional interconnects, ushering in a new era for efficient, scalable high-performance computing.”
“We continue to invest in and develop our HPC and AI technologies to further boost application performance for our customers and help them unlock greater value from their research, engineering, and business initiatives,” said Justin Hotard, senior vice president and general manager, HPC and AI, at HPE. “By partnering with Ayar Labs, we will advance innovation for the HPC and AI market, and leverage their expertise in optical I/O in future generations of HPE Slingshot to deliver unprecedented bandwidth and speed, at lower levels of power and latency, to meet requirements for growing demands in scale and performance.”