InfiniBand and RoCE Connect 71% of All TOP500 Supercomputers

The November edition of the TOP500 supercomputer rankings, released during SC25, reaffirms the pivotal role of InfiniBand and Ethernet RoCE fabrics in advancing global high-performance computing (HPC) and AI-scale infrastructure.

With AI factories, simulation workloads, and data-intensive research accelerating at unprecedented rates, the industry continues to lean heavily on high-performance networking. The latest results highlight how RDMA-based networks are essential elements of the world’s most powerful systems.

  • 362 systems (71%) use InfiniBand or Ethernet RoCE fabrics
  • 276 systems (55%) run on InfiniBand, reinforcing its position as the fabric of choice for large-scale scientific computing, climate modeling, and AI training
  • 35 of the 45 new systems (78%) deploy InfiniBand or Ethernet RoCE, with:
    • 31 new InfiniBand systems
    • 4 new Ethernet RoCE systems

Figure 1. TOP500 Interconnect Trend

The continued preference for accelerated fabrics demonstrates the industry’s commitment to reducing communication bottlenecks, improving real-world application throughput, and enabling efficient scaling across tens of thousands of compute nodes.

InfiniBand Helps Drive Energy-Efficient HPC

Sustainability remains a major theme across HPC and AI system design, and the Green500 list reveals the growing correlation between energy efficiency and high-performance networking.

On the latest list:

  • 7 of the Top 10 Green500 systems run on InfiniBand
  • These include positions 1, 2, 3, 5, 6, 7, and 9

This leadership reinforces a clear trend that energy-efficient supercomputing architectures rely on high-throughput, low-latency, congestion-controlled networks such as InfiniBand.

By reducing data movement overhead and maximizing node-to-node efficiency, InfiniBand helps operators meet both performance and sustainability goals. This is critical for AI factories and exascale-class deployments.

Networking for Exascale, AI Factories, and Next-Generation Research

As scientific research and AI workloads continue to intertwine, the ability to move data efficiently across massive, distributed compute systems becomes an essential architectural pillar.

The November TOP500 list again demonstrates why accelerated networking has become foundational:

  • InfiniBand’s hardware-accelerated transport, adaptive routing, and in-network computing technologies enable higher job throughput and faster time-to-solution.
  • Ethernet RoCE offers an attractive accelerated option for cloud-native environments adopting RDMA at scale.
  • Both fabrics continue to expand across newly deployed systems—from academic research clusters to industrial AI training platforms to national-scale supercomputing programs.

With the majority of the world’s top systems relying on InfiniBand and Ethernet RoCE, the ecosystem enters 2026 with strong alignment around high-efficiency, high-bandwidth networking as a critical enabler of HPC and AI progress.

Brian Sparks

IBTA Marketing Working Group Chair