
Cornelis has introduced its CN6000 networking platform, combining an 800G Omni-Path fabric with a new Ethernet SuperNIC supporting RoCEv2 for AI training, disaggregated inference, and high-performance computing (HPC). The platform includes an Omni-Path SuperNIC, a 48-port Omni-Path switch, and an Ethernet SuperNIC designed for industry-standard switches. Cornelis reports 1.6 billion messages per second and 0.95 microseconds of MPI ping-pong latency for its Omni-Path adapter, versus 1.2 microseconds for NVIDIA InfiniBand XDR at 800G in company-run testing. The approximately 21% latency advantage has not been independently verified.
The Omni-Path SuperNIC doubles the bandwidth of the preceding CN5000 generation to 800 Gb/s and incorporates credit-based lossless delivery, multi-plane and multi-rail operation, and hardware security features including SPDM 1.4 attestation and post-quantum ML-DSA signatures. The companion 1U Omni-Path switch provides 48 ports of 800G connectivity, corresponding to 38.4 Tb/s of summed one-directional port bandwidth, and adaptive routing across paths. Cornelis’ Ethernet SuperNIC brings hardware-accelerated RoCEv2 to standard Ethernet switches for latency-sensitive traffic such as key-value cache transfers in disaggregated inference. Ultra Ethernet support is planned through software updates.
CN6000 supports AMD, Arm-based, and Intel CPUs and accelerators from AMD, NVIDIA and Qualcomm. Cornelis is validating the platform with AMD EPYC 9006 Series processors and next-generation AMD Instinct GPUs, while extending its Intel Xeon collaboration to CN6000. Dell Technologies, Hyve Solutions, Lenovo, NEC, Penguin Solutions and Supermicro are among OEMs and integrators planning validation or support. CN6000 SuperNICs, switches and software are available for evaluation and early deployments, with partner systems expected beginning in Q1 2027. Cornelis plans an SC26 demonstration in Chicago, November 15–20.
“Every minute an accelerator waits on the network is capacity our customers paid for and didn’t use,” said Lisa Spelman, CEO of Cornelis. “We built CN6000 for the communication patterns that make parallel AI and HPC workloads so demanding.”
| Cornelis CN6000 | Key Facts | |
| Bandwidth | • 800 Gb/s per SuperNIC port; twice CN5000’s bandwidth |
| Message rate | • 1.6 billion messages per second, Cornelis-reported |
| MPI latency | • 0.95 µs, versus 1.2 µs for NVIDIA InfiniBand XDR in Cornelis’ internal 8-byte ping-pong testing |
| Omni-Path switch | • 48 × 800G ports in 1U; 38.4 Tb/s summed one-directional port bandwidth; adaptive routing |
| Ethernet | • RoCEv2 SuperNIC works with industry-standard 800G Ethernet switching; Ultra Ethernet planned via software |
| Host and security | • PCIe 6.0; SPDM 1.4 attestation; ML-DSA signatures on Omni-Path SuperNIC |
| Availability | • Evaluations and early deployments now; OEM partner systems targeted for Q1 2027 |
Analysis
CN6000 combines two networking models within one product generation. Omni-Path targets tightly coupled HPC and AI communication, while its Ethernet SuperNIC offers a path into deployments using conventional Ethernet switching. HPC jobs often exchange many small messages, making message rate and communication latency important alongside peak bandwidth. AI training and inference fabrics, meanwhile, mix bulk movement and latency-sensitive communication across accelerators and memory resources.
Cornelis’ 0.95-microsecond figure comes from an 8-byte MPI ping-pong test using Intel MPI, rather than an end-to-end AI training or inference workload. Application outcomes depend on network topology, load, congestion, host integration and communication libraries. The claim should therefore be interpreted as a vendor-reported microbenchmark comparison, not proof of a comparable percentage improvement in overall application performance.
The Ethernet adapter gives Cornelis access to RoCEv2-based deployments, including distributed inference architectures that move KV-cache data between nodes. RDMA performance in practice depends on network design, switch configurations and congestion management. Cornelis has not specified the date or complete feature scope of its Ultra Ethernet software support.
The product introduction follows Cornelis’ September 2026 announcement of $205 million in financing and its Active Compute Fabric roadmap spanning scale-out and scale-up networking. That larger architectural program should be distinguished from the CN6000 products currently offered for evaluation.
| Company Profile | Cornelis Networks | |
| Headquarters | Wayne, Pennsylvania, USA |
| Leadership | Lisa Spelman, CEO |
| Ownership | Privately held |
| Portfolio | CN5000 and CN6000; Omni-Path and RoCEv2 SuperNICs, switches and software; Active Compute Fabric roadmap |
| Technology | Lossless transport, adaptive routing, RDMA, high message-rate networking, scale-out and scale-up fabrics |
| Funding | $205 million announced September 2026 |
| Website | cornelis.com |
We’re tracking AI infrastructure and data center networking developments. Follow ongoing coverage at Converge Digest AI Infrastructure.
