Cornelis Networks, the company that has taken over Intel Omni-Path when Intel divested the technology, now has a new 400Gbps generation. The Cornelis CN5000 brings 400Gbps Omni-Path technology to the AI and HPC markets. At the same time, the company shared a vision for the future with the adoption of UltraEthernet for next-generations of products.
Cornelis CN5000 400Gbps Omni-Path Launched
400Gbps networking has ramped very quickly. A big driver in that has been the AI build-out. Faster networking in AI means that huge numbers of GPUs can send and receive work faster and therefore be utilized at a greater rate. Nothing says ROI like being able to increase the utilization rate of billions of dollars of GPUs.
The new Cornelis CN5000 is designed with the 400G SuperNICs (SuperNIC is also a term that NVIDIA and others have used. Thsoe connect to either 48 port switches or up to 576-port director switches. Cornelis makes the SuperNIC and switch ASIC for the ecosystem.
The director class switch in this generation is not just air cooled, but there is a liquid cooled version as well. Liquid cooled switches is something we have heard multiple large AI shops discuss as a need since with liquid cooled GPU compute nodes, the networking side still generating a lot of air cooled heat means that the networking side is cooled less efficiently. It also has implications for facility design when only some of the infrastructure can be liquid cooled.
Cornelis says that the new class of devices are better suited to newer workloads.
The lossless and zero congestion architecture is a big update from the 100G Omni-Path OPA100 generation.
It appears as though it is launching at ISC 2025 even though we saw the CN5000 series at SC23. GA will happen next quarter.
For some reason the director class switches are always neat to see.
Cornelis UltraEthernet Roadmap
Perhaps the more interesting feature is what happens in the future. Back in the Skylake Xeon days, I was a big proponent that the onboard OPA100 should also do Ethernet, or there should be an Ethernet SKU. At the time, Intel pushed back on this idea since it would hurt its Ethernet adapter business that has largely missed the AI build-out wave. On the Cornelis slide, we get an UltraEthernet reference with the UEC vision.
Starting with the CN6000 series generation in 2026, the SuperNICs will be upgraded to support Ethernet. In 2027 and the 1.6T generation we should get Ultra Ethernet also at the switch level.
It seems that Cornelis is hunting not just NVIDIA, but also Broadcom and Marvell with this move.
Final Words
Almost a decade since I first saw Omni-Path’s 100Gbps generation which was a compute-focused InfiniBand derivative, the network space has heated up considerably.
The new 400Gbps generation matches PCIe Gen5 x16 links in terms of speed, and there is a market for lower-cost NICs. NVIDIA is charging a lot for ConnectX-7 and ConnectX-8. Recently we went into how NVIDIA is also moving to close off the NIC market to anyone other than NVIDIA ConnectX, and the BOM impacts of that move. Still, for many clusters that are not using the ConnectX-8 embedded options, there is a huge opportunity to provide a lower-cost alternative. UltraEthernet will likely be the answer in 2-3 years, but for now, Cornelis has an opportunity.
Cornelis has an opportunity with the new CN5000 series to provide the lower-cost alternative to NVIDIA’s InfiniBand. At the same time, the 2027 CN7000 has us very excited. Hopefully we will get some hands-on time soon.
