Artificial intelligence (AI) workloads are having a real impact on data center network traffic, which analyst firms predict will drive significant technology decisions and investments around the use of InfiniBand or Ethernet to support that traffic.

InfiniBand has been seen as the most performant backend networking technology, which has aligned it with AI demand. However, Ethernet is more widely understood, with both playing a role in AI system networking needs.

Neil MacDonald, EVP and GM for compute, high-performance computing (HPC) and AI at HPE, explained during an investor relations session at last year’s HPE Discover event that AI systems involve multiple different network needs. This includes “fabrics that connect accelerators together,” and then “fabrics that connect sets of accelerators together.” The other two are data center networking that interconnect these fabric connections and “then you have whatever you’re using to get to your data and your storage.”

This diverse set of networking needs leads support for using a diverse set of networking technologies. “I think sometimes when this thread of conversation comes up, it comes up through the lens of what’s the one binary answer,” MacDonald said. “There isn’t going to be one binary answer.”

MacDonald noted that this leaves open the opportunity for both InfiniBand and Ethernet to serve those different needs.

“The biggest shift that’s going to emerge is that InfiniBand, which has been used in some of the highest performance systems in the world, including some of the highest performance AI systems in the world, is something of an alien technology for many enterprises and as a result Ethernet is much more absorbable,” MacDonald said. “But classic Ethernet is not sufficiently performant for these workloads.”

This network use case utilization is becoming increasingly important.

650 Group in a recent report noted that it’s critical to pick the right model to support training and inference in efforts to fully monetize GPU investments. “Suboptimal networking can waste billions of dollars in processor cycles or require expensive restarts,” the firm noted.

Is Ethernet gaining an edge over InfiniBand? That financial consideration could tip the growth balance toward Ethernet in many cases.

650 Group said it expects Ethernet, fueled by growing adoption of 800 Gb/s (800G) and eventually 1.6 Tb/s (1.6T) system capacity, will surpass InfiniBand as “the dominant technology for scale-out” by the end of this year. That ramp will see the overall scale-out networking space more than double by year-end, generating more than $8 billion in revenues.

The analyst firm noted that the market last year was dominated by Celestica, Nvidia, and Arista, with Nvidia’s Ethernet growth notching that vendor as the “fastest-growing vendor in data center switching for 2024.”

Dell’Oro Group noted a similar tilt toward growing Ethernet usage.

“Ethernet is experiencing significant momentum, propelled by supply and demand factors,” the firm noted in a report. “More large-scale AI clusters are now adopting Ethernet as their primary networking fabric.”

Dell’Oro Group, for its part, does sees a longer path toward domination for Ethernet, forecasting the technology will not surpass InfiniBand in revenue share until 2027. It also noted that while its research found Celastica, Huawei, and Nvidia as the market’s largest players last year, it expects other players like Accton, Arista, Cisco, Juniper, and Nokia “to gain ground.”

“We expect the vendor landscape in AI backend networks to remain very dynamic as cloud [service providers] hedge their bets by diversifying their supply on both the compute side and the networking that goes with it,” Dell’Oro Group noted.