Artificial intelligence (AI) workloads and the backend network infrastructure needed to support those applications will expand data center switch spending by 50%, according to research from Dell’Oro Group. But what is still up for debate is whether InfiniBand or Ethernet will emerge as the dominant AI backend network fabric in this soon-to-be $10 billion market.
Spending on AI backend networks currently represents less than 15% of the total data center switch market. By 2027, however, analysts expect AI backend network switches will account for 40% of the data center switch market.
Generative AI (genAI) applications are ushering in a novel era in the age of AI, “standing out for the sheer number of parameters that they have to deal with,” Dell’Oro VP Sameh Boujelbene said.
Some large AI applications are handling trillions of parameters today, and that count will increase tenfold each year. “This rapid growth necessitates the deployment of thousands or even hundreds of thousands of accelerated nodes. Connecting these accelerated nodes in large clusters requires a data center-scale fabric, known as the AI backend network, which differs from the traditional frontend network used mostly to connect general purpose servers,” Boujelbene said.
The firm analyzed different AI backend networks built by major cloud service providers like Google, Amazon Web Services, Microsoft, Meta, Alibaba, Tencent, ByteDance and Baidu. Based on the number of high-end accelerated servers currently deployed, analysts determined that Microsoft’s “is by far the largest, followed by Google and then Meta,” Boujelbene told SDxCentral.
Ethernet vs. InfiniBand
InfiniBand and Ethernet remain in tense competition as manufacturers on each side seek AI backend network market dominance. Dell’Oro anticipates Infiniband will maintain its current lead, but Ethernet should also see substantial gains in adoption – Boujelbene said she expects the latter to reach a 20% share of the market by 2027.
Despite market forecasts, a crucial question remains: “What is the most suitable fabric that can scale to hundreds of thousands – and potentially millions – of accelerated nodes while ensuring the lowest job completion time?”
That debate will likely continue raging on.
While some might argue that Ethernet is at least one speed generation ahead of InfiniBand, network speed isn’t all that matters. InfiniBand and Ethernet handle congestion control and adaptive routing in different ways, Boujelbene said, and those factors can help enterprises determine which fabric is most appropriate for their needs.
“InfiniBand used to be ahead of Ethernet in terms of performance in AI backend networks, but we are seeing significant improvement on Ethernet to close the gap,” she said.