Cornelis Networks says its CN6000 800G networking family is shipping now, with systems from OEM and integration partners expected from Q1 2027. The launch adds an Ethernet option alongside Omni-Path, giving AI and high-performance computing operators a choice of fabric, not just another impressive number on a product sheet.
Cornelis Networks Watch analysis
What happened
CN6000 includes Omni-Path and Ethernet SuperNICs, built for PCIe 6.0 platforms. Cornelis says its Ethernet SuperNIC supports RoCEv2 on standards-based switches, with Ultra Ethernet available through early access. Its Omni-Path design offers two independent 400 Gb/s paths per node, which can run together for up to 800 Gb/s or keep compute and storage traffic on separate planes.
In tests conducted by Cornelis, the Omni-Path SuperNIC recorded 0.97 microseconds of MPI latency, compared with 1.23 microseconds for an 800G XDR InfiniBand configuration. At 128 processes per node, it reached 809 million messages per second, about three times the comparison result. These are vendor-run results, and Cornelis notes that performance varies with workload and configuration. The company says CN6000 can join existing CN5000 Omni-Path fabrics using the same software and host driver. Read Cornelis’s CN6000 overview.
Our top picks
- Two fabrics in one family
Operators can choose Omni-Path or an Ethernet SuperNIC supporting RoCEv2, rather than being pushed towards one networking fabric. - Independent network planes
The Omni-Path design can separate workloads across two 400 Gb/s paths, or use both for up to 800 Gb/s per node. - A measurable performance claim
Cornelis reports 0.97-microsecond MPI latency and 809 million messages per second at 128 processes per node in its comparison tests. - A route from existing CN5000 deployments
The company says CN5000 and CN6000 share software and a host driver, allowing them to operate in one Omni-Path fabric. - Hardware security features
Cornelis says every card includes a hardware root of trust, attestation and post-quantum firmware-signing support.
Why it matters
In large AI and HPC clusters, networking affects how quickly accelerators can exchange data and finish work. The product’s practical pitch is flexibility: an Ethernet option for teams already using standards-based switches, and an Omni-Path route for operators prioritising latency and message rate. The dual-plane design also aims to keep a failure on one path from stopping the other.
The benchmark figures are useful evidence of what Cornelis says it can achieve, not a promise that every workload will see the same result. Buyers will want independent testing and details for their own cluster before treating a vendor comparison as a purchasing verdict.
Our read
This is a substantial infrastructure launch, with a clearer choice of fabrics and a stated path for existing CN5000 customers to grow. The most interesting claim is not simply “800G”, but the combination of separate network planes and compatibility across generations. Operators should compare the numbers with their own workloads, and ask how the Ethernet and Omni-Path options fit their current switches and software.
What to watch
- Whether OEM and system-partner deployments arrive on the stated Q1 2027 schedule.
- Independent tests of latency and message rates against comparable configurations.
- Which existing switches and platforms qualify for the Ethernet SuperNIC.
- How CN6000 performs in real mixed AI, storage and checkpoint workloads.
Discussion spark: For a new AI cluster, would you choose an open Ethernet fabric or a dedicated low-latency network, and what evidence would decide it?
Sources and evidence
- Source update (17 November 2026, 06:00 UTC)
Independent WittyWires tracker for public updates about Cornelis Networks. Not affiliated with or endorsed by Cornelis Networks; this is not an official account.