The Open Compute Project (OCP) has unveiled a new initiative called Ethernet for Scale-Up Networking (ESUN), aiming to establish open standards for high-performance connections in artificial intelligence (AI) clusters. Major players in the tech industry, including Meta, Nvidia, OpenAI, AMD, and Cisco, are collaborating to examine how Ethernet can compete with established interconnects like InfiniBand in large-scale data centers. Other notable companies involved in this project include Arista, ARM, Broadcom, HPE Networking, Marvell, Microsoft, and Oracle.
The announcement comes as InfiniBand has maintained a stronghold on the AI networking market, controlling approximately 80% of the infrastructure connecting graphics processing units (GPUs) and accelerators. Nevertheless, the ESUN group believes that the maturity, cost-effectiveness, and interoperability of Ethernet present a viable alternative for scaling AI clusters. Unlike proprietary systems, Ethernet’s widespread familiarity among engineers may simplify the management of extensive AI workloads.
Supporters of the initiative argue that adopting Ethernet as an open standard will enable operators to expand their infrastructure while reducing costs. The ESUN initiative builds on prior work under the SUE-Transport (SUE-T) program, which investigated Ethernet transport for multi-processor systems. Participants in ESUN will convene regularly to define standards related to switch behavior, including protocol headers, error handling, and lossless data transfer.
Collaborative Efforts to Redefine Networking Standards
The group will also research how network design influences load balancing and memory ordering in GPU-based systems. They plan to coordinate with the Ultra Ethernet Consortium and the IEEE 802.3 standards body to ensure consistency across the broader Ethernet ecosystem. Several firms have already developed Ethernet-based solutions aimed at AI scalability. For instance, Broadcom’s Tomahawk Ultra switch can handle up to 77 billion packets per second, while Nvidia’s Spectrum-X platform integrates Ethernet with acceleration hardware for AI clusters.
As a co-founder of the OCP in 2011, Meta views the ESUN initiative as a natural progression of its commitment to open hardware within data centers. Nonetheless, industry observers caution that for Ethernet to replace the established InfiniBand networks, it must demonstrate superior performance capabilities under demanding AI workloads, particularly regarding latency and reliability.
The success of ESUN hinges on its ability to balance openness with high performance. Advocates envision a future where AI systems operate on interoperable hardware utilizing standardized Ethernet technologies. Yet, given the scale and sensitivity of AI infrastructure, it remains uncertain whether the industry will decisively pivot away from proprietary interconnects.
Ultimately, the ESUN initiative represents an ambitious effort to redefine networking standards in the AI sector. The ability of Ethernet to match or exceed the performance of InfiniBand will be closely monitored as this collaborative endeavor unfolds.
