Skip to content
CN6000 800G SUPERNICS AND SWITCH Omni-Path® or Ethernet

Cornelis CN6000

Performance where it matters.
Flexibility where you need it.

  • Request a Design Consult Arrow - Right
  • Download the product brief Download
  • Overview

    What is CN6000?

    AI and HPC returns depend on keeping expensive accelerators busy. Cornelis® CN6000™ is the industry’s first 800G SuperNIC family that includes both Omni-Path® and Ethernet connectivity options, so you can choose the fabric that fits your workload and network.

    The CN6000 Omni-Path SuperNIC delivers 1.6 billion messages per second and sub-microsecond MPI latency on a lossless fabric that scales past 500,000 endpoints.¹ Paired with the 48-port CN6000 Omni-Path Switch, it helps training runs and simulations finish sooner. Applications carry forward from CN5000 to CN6000 and on to CN7000 with a fully compatible software stack.

    For Ethernet-first AI, the CN6000 Ethernet SuperNIC brings standards-based RoCEv2 to fine-tuning and inference, and it is Ultra Ethernet compliant (Preview). Its lightweight queue pairs handle the bursty inter-GPU traffic of distributed fine-tuning.

    Key takeaways:

    • More work from every XPU: 1.6 billion messages per second and 800 Gb/s per Omni-Path SuperNIC cut the time accelerators spend waiting on the network.

    • Bigger fabrics, smaller fault domains: Two independent 400 Gb/s data paths, matched by a dual-plane 800G switch, support multi-plane and multi-rail designs.

    • Serve more tokens and accelerate fine-tuning:  The Omni-Path SuperNIC moves disaggregated inference traffic with sub-microsecond latency. The Ethernet SuperNIC brings RoCEv2 to fine-tuning and inference on standards-based switches and is Ultra Ethernet compliant (Preview).

    CN6000 at a glance: four headline figures for the Omni-Path and Ethernet SuperNICs.

    • 800GBandwidth per SuperNIC, over PCIe 6.0 x16
    • 1.6BMessages per second, bidirectional, Omni-Path SuperNIC¹
    • <1 µsMPI latency, Omni-Path SuperNIC¹

    ¹ Results vary with configuration, workload and topology.

    Benefits

    Turn communication time into compute time

    Accelerators are the largest line item in an AI or HPC cluster. CN6000 helps you get more from each one, grow the fabric safely, and serve inference on the network you choose.

    Get more from every XPU

    Faster time to results and higher accelerator utilization

    Collectives, small messages and sync stalls leave GPUs idle. The CN6000 Omni-Path SuperNIC pairs the industry's highest message rate with 800 Gb/s of bandwidth. Communication keeps pace with compute, so training runs and simulations finish sooner.

    Build bigger, more resilient fabrics

    800G multi-plane and multi-rail designs that scale and keep jobs running

    Two independent 400 Gb/s data paths let each node connect to two switch planes. You get up to 800 Gb/s per node while each plane keeps the radix and low hop count of a 400G fabric. A fault in one plane stays in that plane, so long jobs keep running.

    Serve more tokens and accelerate fine-tuning

    Omni-Path or RoCEv2 for inference; RoCEv2 for fine-tuning on Ethernet

    KV cache handoffs and attention-FFN exchanges sit on the critical path of disaggregated inference. The CN6000 Omni-Path SuperNIC moves that traffic with sub-microsecond latency, which helps lower time to first token and cost per token. For Ethernet-first teams, the CN6000 Ethernet SuperNIC brings RoCEv2 to inference and fine-tuning on standard switches.

    Highlighted Feature

    800G per node with the scale of a 400G fabric

    Faster switch ports usually cost you radix, which adds tiers, hops, optics and cost as clusters grow. A single plane is also a single failure domain. CN6000 takes a different path.

    Each CN6000 Omni-Path SuperNIC port carries two independent 400 Gb/s links. The CN6000 Omni-Path Switch is built the same way, with two independent 48-port 400G planes in one 1U switch. With one straight cable, every node lands on both planes at up to 800 Gb/s, while each plane stays a 400G fabric with full radix and a low hop count.

  • Talk to us about your fabric design Arrow - Right
  • cn6000 omni path switch 48x800g qsfp dd 1u
    • 800 Gb/s per node with multi-rail

      The OPX libfabric provider drives both data paths at once, so one job can use the full bandwidth of both planes.

    • Faults stay contained

      A switch, link or optics failure in one plane leaves the other plane running, so the cluster keeps working while you service it.

    • Rail-optimized AI clusters

      Connect GPU number n in every node to rail n, so collectives between same-index GPUs stay on one rail.

    • Fewer tiers at scale

      Each plane keeps 400G switch radix and a low hop count, so large clusters need fewer tiers, hops and optics.

    • Separate traffic by role

      Put compute on one plane and storage or checkpoint on the other, so they stop competing for the same paths.

    • Same fabric, more scale

      CN6000 SuperNICs also drop into existing CN5000 switching, so your teams grow bandwidth without learning a new protocol.

    Features

    Two SuperNICs, your choice of fabric

    CN6000 comes as two SuperNICs, each built for the fabric you choose. The Omni-Path SuperNIC pairs with the CN6000 Omni-Path Switch for an end-to-end Omni-Path fabric. The Ethernet SuperNIC connects to standards-based Ethernet switches. OPX software runs the Omni-Path fabric.

    OMNI-PATHOmni-Path SuperNIC, switch and software

    For HPC and performance-intensive AI, including disaggregated inference. The Omni-Path SuperNIC and the CN6000 Omni-Path Switch form an end-to-end 800G fabric, natively designed for multi-plane, and OPX software runs it.

    CN6000 OMNI-PATH® SuperNIC Preview shot of the CN6000 SuperNIC 1.6B msg/s bidirectional message rate

    Omni-Path SuperNIC

    PDF

    The CN6000 Omni-Path SuperNIC delivers 800 Gb/s over two independent 400 Gb/s data paths on a low-profile PCIe 6.0 x16 adapter. Credit-based flow control, link-level retry and fine-grained adaptive routing keep the fabric lossless as it grows past 500,000 endpoints. Sub-microsecond MPI latency helps large-scale AI training and HPC codes finish sooner.

    Learn more about the Omni-Path SuperNIC Arrow - Right
    CN6000 OMNI-PATH® Switch Preview shot of the CN6000 Switch 76.8 Tb/s full-duplex switching bandwidth

    Omni-Path Switch

    PDF

    The CN6000 Switch delivers 48 ports of 800 Gb/s connectivity, or 96 ports at 400 Gb/s, in a 1U liquid-cooled platform. It is natively designed for multi-plane fabrics: two independent 400G planes let one cable per node reach both. Fine-grained adaptive routing, credit-based flow control and switch latency as low as 105 ns keep AI and HPC traffic moving under load.

    Learn more about the Omni-Path Switch Arrow - Right
    CORNELIS OPX SOFTWARE SUITE Preview shot of OPX Software One driver shared by CN5000 and CN6000

    OPX Software for
    Omni-Path

    Cornelis OPX is an open software stack that spans host drivers, fabric services and switch management. The OPX libfabric provider drives both data paths for up to 800 Gb/s per node with multi-rail, and it uses the same provider and host driver as CN5000, so both generations share one fabric. MPI, OpenSHMEM, NCCL and RCCL run unchanged.

    Learn more about OPX Arrow - Right

    ETHERNETEthernet SuperNIC

    For Ethernet-first AI, including fine-tuning and inference on Ethernet networks. The Ethernet SuperNIC connects to the standards-based Ethernet switches you choose.

    CN6000 ETHERNET SuperNIC Preview shot of the Ethernet SuperNIC

    Ethernet SuperNIC

    The CN6000 Ethernet SuperNIC brings hardware-accelerated RoCEv2 to standards-based Ethernet switches and is Ultra Ethernet compliant (Preview). Lightweight queue pairs and an RiF table that tracks millions of in-flight operations handle the many small transfers of inference and the bursty traffic of fine-tuning. Priority Flow Control with 8 priorities keeps traffic lossless, and IEEE 1588 PTP fits the Ethernet network you already run.

    Learn more about the Ethernet SuperNIC Arrow - Right
    Advancements

    Technology that moves the fabric forward

    CN6000 carries the proven Omni-Path architecture into a new generation. Six advances shape what it can do for your cluster.

    Double the bandwidth

    Each SuperNIC moves 800 Gb/s over a PCIe 6.0 x16 host interface, twice the bandwidth of the previous generation. Nodes keep pace with faster accelerators.

    Two independent data paths

    Two 400 Gb/s paths can each reach their own switch plane, and the CN6000 Switch carries two independent planes in one 1U box. That enables multi-plane and multi-rail designs, and it contains a fault to one plane.

    Higher message rate

    Optimized send and receive buffer management drives 1.6 billion messages per second on the Omni-Path SuperNIC. MPI, SHMEM and collectives spend less time on the wire.

    An Ethernet SuperNIC built on Omni-Path design

    A new Ethernet adaptation layer brings Omni-Path design to an Ethernet SuperNIC. Choose the Omni-Path SuperNIC for HPC and performance-intensive AI, or the Ethernet SuperNIC for Ethernet-first AI.

    RoCEv2 for inference and fine-tuning

    Hardware-driven RDMA, lightweight queue pairs and an in-flight tracking table handle the many small transfers of disaggregated inference and the bursty traffic of fine-tuning.

    Post-quantum security

    A Root of Trust secured with post-quantum cryptography supports SPDM 1.4 with ML-DSA-87 and ECDSA P-384. Firmware and identity stay protected as threats evolve.

    In Their Words

    Industry leaders on CN6000

    Silicon and systems partners see CN6000 as a way to raise performance, scale out, and bring standards-based Ethernet to AI and HPC.

    GET MORE FROM EVERY XPU

    "As our compute footprint has grown, maintaining high utilization across our accelerators has become critical. The CN6000 promises to provide the consistent, low-latency fabric necessary to keep our training and simulation workloads running efficiently into the 800G infrastructure upgrade cycle."

    Alex Lesser

    Vice President, Sales and Marketing, PSSC Labs

    BUILD BIGGER, MORE RESILIENT FABRICS

    "We believe open, standards-based innovation is key to unlocking scalable, high-performance infrastructure for the future of AI and HPC. Industry collaborations like the Ultra Ethernet Consortium help ensure a diverse and interoperable ecosystem, empowering customers with choice and flexibility to meet evolving compute demands."

    Robert Hormuth

    Corporate Vice President, Architecture and Strategy, AMD

    SERVE MORE TOKENS ON STANDARD ETHERNET

    "Lenovo is focused on building HPC and AI-optimized systems that scale. We expect Ultra Ethernet to be transformative for the AI market, and are looking forward to the CN6000 and enabling our customers who want no-compromise high performance with Ethernet-based infrastructures."

    Scott Tease

    VP and GM, HPC and AI, Lenovo

    Source: Cornelis CN6000 launch announcement, November 18, 2025

    Ecosystem

    Systems, platform, and channel partners

    Omni-Path runs in production at Lawrence Livermore National Laboratory on the Lynx system and at the Texas Advanced Computing Center on Stampede3, both with CN5000. CN6000 brings the same fabric to 800G with support from server, silicon, software and channel partners. The logo grid below lists those partners.

    Woman in high-rise office looking out her window thinking with her hand to her chin
    • AMD logo
    • Lenovo logo
    • Atipa logo

    Reference architecture: CN6000 with AMD EPYC and AMD Instinct

    A validated design that pairs CN6000 with AMD EPYC (Venice) processors and AMD Instinct MI400 series accelerators.

  • Read the reference architecture Arrow - Right
  • Pricing And Availability

    How to get CN6000

    Server OEMs and channel partners sell CN6000 in configured systems. Pricing depends on your configuration, so there is no single list price.

    How CN6000 is sold

    OEMs and system integrators deliver CN6000 as part of configured AI and HPC systems. Federal buyers can work through approved integrators.

    What shapes a quote

    • Scale: Node count and endpoint targets.

    • SuperNIC: Omni-Path SuperNIC or Ethernet SuperNIC.

    • Fabric design: Single plane, or multi-plane and multi-rail.

    • Switching: CN6000 Switch, existing CN5000 switching, or third-party Ethernet switches.

    • Cooling: Air or conduction for SuperNICs; liquid for the switch.

    Availability

    Ask Cornelis or a partner for current availability. Lead times vary by configuration.

    Key Specifications

    CN6000 at a glance

    The lists below give the main SuperNIC and switch specifications. See each product brief for full details.

    CN6000 SuperNICs
    Bandwidth 800 Gb/s; 1.6 Tb/s bidirectional
    Ports Dual x4-lane QSFP-DD, 112 Gb/s per lane
    Host interface PCIe 6.0 x16
    SuperNICs Omni-Path SuperNIC; Ethernet SuperNIC (RoCEv2; Ultra Ethernet compliant (Preview))
    Message rate 1.6B messages/s, bidirectional (Omni-Path SuperNIC)
    Latency Sub-1 µs MPI (Omni-Path SuperNIC)
    Virtual lanes and MTU 1 to 4 data VLs plus 1 management VL; MTU 2, 4, 8 or 10 KB per VL (Omni-Path SuperNIC)
    Security PQC Root of Trust; SPDM 1.4 (ML-DSA-87, ECDSA P-384)
    Management SR-IOV; PLDM, MCTP and Redfish (RDE)
    Form factor PCIe low profile, 3.12 x 6.6 in
    Cooling Air-cooled or conduction-cooled
    Typical power 42.4 W (DAC); 57.5 W (AOC)
    CN6000 Omni-Path Switch
    Bandwidth 76.8 Tb/s full duplex
    Ports 48 x 800G QSFP-DD, or 96 x 400G
    Architecture Two independent 48-port 400G planes
    Routing Fine-grained adaptive and static dispersive routing
    Flow control Credit-based, incast-aware; link-level retransmission
    Switch latency As low as 105 ns
    Virtual lanes 1 to 8 data VLs plus 1 management VL
    Security Secure boot and protected key storage
    Management OpenBMC with Redfish and PLDM; in-band and out-of-band; CLI
    Form factor 1U; 17.2 x 26.1 in
    Cooling and power Liquid-cooled; 240 to 277 VAC or 48 VDC; N+N hot-swap supplies
    Typical power 856 W (all DAC); 1,663 W (all AOC)
    Open Software

    Open software, on the path to an open fabric

    Cornelis is moving toward an open fabric that spans Omni-Path and standards-based Ethernet. CN6000 is a step on that path: Cornelis now offers an Ethernet SuperNIC beside the Omni-Path SuperNIC, on open-source software your teams already use.

    • Built on libfabric: OPX runs on libfabric (OFI), the open interface that Cornelis originally developed. libfabric · Cornelis OPX

    • Open standards: The Ethernet SuperNIC runs standards-based RoCEv2 and is Ultra Ethernet compliant (Preview), which reflects our commitment to open standards. Ultra Ethernet Consortium

    • Frameworks you already run: MPI, OpenSHMEM, NCCL, RCCL, UCX and PyTorch.

    • Any accelerator: GPUDirect for NVIDIA and ROCm RDMA AMD accelerators.

    • A path forward: CN5000 and CN6000 share one driver, and applications carry forward to CN7000 with a fully compatible software stack.

    • Operating systems: The Omni-Path SuperNIC runs on the Linux distributions listed on the OPX page. The Ethernet SuperNIC supports Red Hat Enterprise Linux (RHEL), with other distributions on request. OPX page

    YOUR QUESTIONS, ANSWERED

    Frequently asked questions

    Everything about the CN5000 fabric in one place — architecture, positioning against InfiniBand and Ethernet, roadmap and procurement.

    • What is the Cornelis CN6000? CN6000 is an 800G SuperNIC family for AI and HPC with two SuperNICs: the Omni-Path SuperNIC and the Ethernet SuperNIC (RoCEv2, Ultra Ethernet compliant (Preview)). The Omni-Path SuperNIC delivers 1.6 billion messages per second and sub-microsecond MPI latency on a lossless fabric. The CN6000 Omni-Path Switch provides 48 ports at 800G for full Omni-Path fabrics.
    • When should I choose the Omni-Path SuperNIC or the Ethernet SuperNIC? Choose the Omni-Path SuperNIC for HPC codes, large-scale AI training and disaggregated inference, where latency and message rate set the pace. Choose the Ethernet SuperNIC when your design calls for standards-based RoCEv2 on Ethernet switches, such as fine-tuning or inference on an Ethernet network.
    • Can I fine-tune models on the CN6000 Ethernet SuperNIC? Yes. RoCEv2 lightweight queue pairs handle the massive, bursty inter-GPU communication that fine-tuning demands across distributed clusters. NVMe-oF over RoCEv2, for fast access to shared flash, is on the roadmap.
    • What's the best network setup for splitting prefill and decode across separate GPU pools? Use a lossless, low-latency, high-message-rate fabric between the prefill and decode pools. Disaggregated inference moves KV cache from prefill to decode nodes for each request, and it exchanges small activations between attention and FFN layers for each token. Latency and message rate sit on that path, so they shape time to first token and cost per token. The CN6000 Omni-Path SuperNIC handles this traffic with sub-microsecond latency, 1.6 billion messages per second, and credit-based flow control that keeps the fabric lossless. Teams that standardize on Ethernet can use the CN6000 Ethernet SuperNIC, which runs hardware-accelerated RoCEv2.
    • Does CN6000 support rail-optimized AI training clusters? Yes. In a rail-optimized design, GPU number n in every node connects to rail n, so collectives between same-index GPUs stay on one rail. CN6000 SuperNICs can build these rail fabrics on Omni-Path or Ethernet, with a separate Ethernet network for control-plane and management traffic.
    • Does CN6000 work with my existing CN5000 fabric? Yes. CN6000 uses the same OPX provider and host driver as CN5000, so CN5000 and CN6000 SuperNICs can share one Omni-Path fabric. Applications built for CN5000 run on CN6000 with the same software stack.
    • How does CN6000 fit the Cornelis product roadmap? CN5000 is the shipping 400G Omni-Path generation. CN6000 takes the fabric to 800G and adds an Ethernet SuperNIC, a step toward an open fabric that spans Omni-Path and standards-based Ethernet. Applications carry forward from CN5000 to CN6000 and on to CN7000 with a fully compatible software stack. CN7000, the next generation, extends the roadmap to Active Compute Fabric and in-network compute.
    • How is CN6000 sold? Server OEMs, system integrators and channel partners sell CN6000 in configured systems. Quotes depend on scale, SuperNIC type, switching, fabric design and cooling.
    • Does Cornelis make both the switch and the NICs, or do I buy them separately? Cornelis makes both. CN6000 includes the Omni-Path SuperNIC, the Ethernet SuperNIC and the CN6000 Omni-Path Switch, and open-source OPX software runs the Omni-Path fabric from host drivers to switch management. OEMs, system integrators and channel partners deliver them together in configured systems. The Ethernet SuperNIC also works with the standards-based Ethernet switches you choose.
    • Which switches scale a GPU training cluster past 500 nodes? The CN6000 Omni-Path Switch puts 48 ports at 800G, or 96 at 400G, in one 1U switch, on a lossless fabric that scales to more than 500,000 endpoints. Its two independent 400G planes keep 400G radix, so large clusters need fewer tiers, hops and optics. Fine-grained adaptive routing, credit-based flow control and switch latency as low as 105 ns keep traffic moving under load.

    Ready to evaluate CN6000?

    Tell us about your cluster, workloads and scale targets. We will help you pick the right SuperNIC and fabric design.

  • Request a design consult Arrow - Right
  • Omni-Path fabrics scale fluidly to 500k+ endpoints