A coalition of the biggest names in tech, including Meta, Nvidia, and OpenAI, is joining forces to make the common networking technology Ethernet the new standard for connecting massive artificial intelligence clusters. The new initiative, called Ethernet for Scale-Up Networking (ESUN), aims to challenge the dominance of the current high-speed networking technology, InfiniBand.
For years, InfiniBand has been the go-to choice for connecting the powerful GPUs and accelerators that run AI, controlling about 80% of the market. But the ESUN group believes that Ethernet, which is more mature, cheaper, and more widely understood by engineers, is a better long-term solution for building the massive data centers of the future.
The goal is to create a new set of open standards for using Ethernet in high-performance AI systems. Supporters argue that this will allow companies to build and expand their AI infrastructure at a lower cost and with less complexity. Unlike proprietary systems like InfiniBand, Ethernet’s widespread familiarity would make it easier to manage the huge, complex workloads of modern AI.
The group, which also includes AMD, Cisco, and Microsoft, will work together to define standards for everything from switch behavior to how data is transferred without any loss.
While some companies like Nvidia and Broadcom have already developed their own Ethernet-based products for AI, this new, collaborative effort is a major step toward creating a truly open and interoperable ecosystem. The big question is whether Ethernet can prove it can match the performance of InfiniBand in the most demanding AI applications, where every millisecond of latency counts.
