Cornelis has introduced the CN6000 family, an 800G networking platform aimed at AI and high-performance computing (HPC) clusters that are running into communication bottlenecks as accelerator counts climb. The lineup includes two 800G SuperNIC options—Omni-Path and Ethernet (RoCEv2)—plus a 1U Omni-Path switch and supporting software.
On the Omni-Path side, the CN6000 Omni-Path SuperNIC is rated at 1.6 billion messages per second at 800G. Cornelis also reports sub-microsecond MPI latency and claims a 21% lower 8-byte MPI ping-pong latency than NVIDIA InfiniBand XDR at 800 Gb/s (0.95 µs vs. 1.2 µs), based on internal testing using Intel MPI. For fabric design, the SuperNIC supports multi-plane and multi-rail configurations, which operators use to segment traffic and improve resiliency under link failures.
The Ethernet version is the CN6000 Ethernet SuperNIC, which brings hardware-accelerated RoCEv2 into 800G Ethernet environments and is intended to work with “industry-standard” Ethernet switches. Cornelis calls out latency-sensitive AI traffic such as KV-cache transfers in disaggregated inference as a target use case. Ultra Ethernet support is planned “through software updates.”
Cornelis is also shipping a CN6000 Omni-Path switch: a 1U system with 48 ports of 800G. In CN6000 Omni-Path fabrics, the company says adaptive routing is used to spread traffic across available paths to reduce congestion as the fabric scales, and that the switch supports native multi-plane deployments.
Security features called out for the Omni-Path SuperNIC include a dedicated hardware Root of Trust with SPDM 1.4 attestation and post-quantum (ML-DSA) signatures, intended to verify firmware before the SuperNIC joins the system. On the software side, Cornelis lists support for NCCL, RCCL, libfabric, MPI, and OpenSHMEM on Omni-Path, using the same software and host driver as CN5000.
For data center engineers, the practical question is whether the fabric can keep expensive accelerators busy under real workloads, especially when those workloads generate lots of small, latency-sensitive exchanges rather than just bulk throughput. CN6000 is explicitly tuned for that message-rate problem, and the Ethernet SuperNIC is a reminder that “Ethernet-first” AI clusters still want HPC-style latency characteristics.
Cornelis said CN6000 SuperNICs, switches, and software are available “today for evaluation and early deployments,” with OEM and system partner systems expected beginning in Q1 2027. OEM and system partners planning validation and support include Dell Technologies, Hyve Solutions, Lenovo, NEC, Penguin Solutions, and Supermicro. Lisa Spelman, CEO of Cornelis, framed the goal bluntly: “Every minute an accelerator waits on the network is capacity our customers paid for and didn’t use.”
Source: Cornelis


















