Cornelis Networks

In this vendor showdown presentation, Philip Murphy of Cornelis Networks argues that the next step function in AI and HPC performance will come from the network rather than accelerators alone. Building on the company's Omni-Path heritage from Intel, he presents a roadmap in which the CN-5000 and CN-6000 deliver end-to-end adapters and switches, with the CN-6000 beginning a transition to open standards by supporting Ethernet, Ultra Ethernet, and RoCE. The CN-7000 completes that transition, adopting Ultra Ethernet and multi-path resiliency for scale-out and embracing UALink and related standards for scale-up, while the CN-8000 quadruples switch bandwidth and introduces co-packaged optics. The centerpiece concept is a distributed compute fabric that moves computation into the network, targeting 14 to 28 percent recovered GPU efficiency in the first generation given that real-world GPU utilization often sits at or below 50 percent. Murphy cites benchmarked results of 32 percent lower latency and two and a half times the message rate of comparable InfiniBand solutions. In the question round he contrasts HPC traffic, which is latency- and message-rate-sensitive with fine-grained exchanges, against AI's large elephant flows and in-network collectives, asserting that one fabric can serve both. He closes by positioning Cornelis, together with partners, as an open-standards alternative to Nvidia's vertically integrated InfiniBand ecosystem.