CN6000 800G Ethernet SuperNIC
Accelerate fine-tuning and serve more tokens on standard Ethernet.
What is CN6000?
The CN6000 800G Ethernet SuperNIC brings low-latency, hardware-driven RoCEv2 to standards-based 400G and 800G switches, and it is Ultra Ethernet compliant (Preview). It is built for Ethernet-first AI: fine-tuning, where many GPUs exchange bursty traffic, and inference on Ethernet networks, where KV cache transfers help set time to first token and cost per token.
New to CN6000? The family page covers both SuperNICs, partner perspectives and how to buy.
Key takeaways:
Hardware-driven RoCEv2: RDMA offload through standard Verbs (rdma-core).
Built for gigascale: An in-flight tracking table maintains millions of concurrent operations as clusters grow.
Fine-tuning ready: Lightweight queue pairs handle bursty inter-GPU traffic.
No forklift upgrade: Connects to the standards-based 400G and 800G Ethernet switches you already run.
Ethernet at a glance: four headline figures for this SuperNIC.
- Two 400 Gb/sData paths
- MillionsConcurrent operations in the in-flight table
- 1.6 Tb/sBidirectional bandwidth (800 Gb/s each way)
- PCIe 6.0x16 host interface to match 800 Gb/s
The network AI and HPC has been waiting for
For Ethernet-first teams, the CN6000 Ethernet SuperNIC brings hardware-accelerated RoCEv2 to inference and fine-tuning on standard switches, with post-quantum security and standard management built in.
Performance at gigawatt-AI scale
In gigawatt-class AI, collective and inference traffic can overwhelm traditional Ethernet. This SuperNIC runs hardware-accelerated RoCEv2 at 800 Gb/s per port.
Build bigger, more resilient fabrics
Hardware-driven RDMA, lightweight queue pairs and a RoCEv2 in-flight (RiF) tracking table maintain millions of concurrent operations. KV cache handoffs finish faster, which helps raise GPU utilization and token yield per dollar.
Serve more tokens and accelerate fine-tuning
The SuperNIC works with standards-based Ethernet switches and endpoints. Priority Flow Control with 8 priorities keeps traffic lossless, IEEE 1588 PTP fits the network you run, and GPUDirect moves data straight to NVIDIA and AMD accelerators.
Fine-tuning on standard Ethernet
RoCEv2 lightweight queue pairs handle the massive, bursty inter-GPU communication across distributed clusters that fine-tuning demands. A new Ethernet adaptation layer brings proven Cornelis pipeline design to this SuperNIC.
Secured by post-quantum cryptography
A Root of Trust with post-quantum cryptography uses SPDM 1.4 with ML-DSA-87 and ECDSA P-384. It helps protect firmware and device identity against current and future cryptographic threats.
Open and built for every environment
The SuperNIC is Ultra Ethernet compliant (Preview), which reflects our commitment to open standards. PCIe 6.0 x16, SR-IOV, and air- or conduction-cooled versions fit the servers you run.
Where Ethernet fits
Choose the CN6000 Ethernet SuperNIC when your design calls for standards-based RoCEv2 on Ethernet switches, such as fine-tuning or inference on an Ethernet network.
AI solutions
Disaggregated inference, fine-tuning and agentic AI on standard Ethernet.
CN6000 family
Both SuperNICs and the Omni-Path Switch, plus partner perspectives and how to buy.
Which CN6000 SuperNIC fits your cluster
CN6000 is an 800G SuperNIC family with two connectivity options. Both SuperNICs provide 800 Gb/s over two 400 Gb/s data paths on a low-profile PCIe 6.0 x16 card. The fabric you run decides which one fits.
CN6000 Ethernet SuperNICThis Page
- Best Fit
- Fine-tuning and inference on Ethernet networks
CN6000 Ethernet SuperNIC
- Best Fit
- HPC and performance-intensive AI, including disaggregated inference
Description
CN6000 Ethernet SuperNICThis Page
- Protocol
- RoCEv2; Ultra Ethernet compliant (Preview)
CN6000 Ethernet SuperNIC
- Protocol
- Omni-Path
Description
CN6000 Ethernet SuperNICThis Page
- Fabric
- Connects to the standards-based Ethernet switches you choose
CN6000 Ethernet SuperNIC
- Fabric
- The Omni-Path SuperNIC and the CN6000 Omni-Path Switch form an end-to-end 800G fabric, run by OPX software
Description
CN6000 Ethernet SuperNICThis Page
- Switches
- Standards-based Ethernet switches
CN6000 Ethernet SuperNIC
- Switches
- CN6000 Omni-Path Switch
Description
CN6000 Ethernet SuperNICThis Page
- Lossless delivery
- Priority Flow Control, 8 priorities
CN6000 Ethernet SuperNIC
- Lossless delivery
- Credit-based flow control
Description
CN6000 Ethernet SuperNICThis Page
- Software
- Compatible with PyTorch, NCCL, RCCL, UCX, MPI and OpenFabrics
CN6000 Ethernet SuperNIC
- Software
- OPX, which runs on libfabric (OFI)
Description
CN6000 Ethernet SuperNICThis Page
- Existing deployments
- Interoperates with standards-based Ethernet switches and endpoints
CN6000 Ethernet SuperNIC
- Existing deployments
- CN5000 and CN6000 share one driver, and applications carry forward
Description
RoCEv2 in hardware, tuned for inference
The SuperNIC handles RDMA in hardware and tracks millions of operations in flight. It moves data straight to NVIDIA and AMD accelerators, and a post-quantum Root of Trust protects its firmware.
| Hardware-driven RoCEv2RDMA offload through standard Verbs (rdma-core). | Lightweight queue pairsLow per-connection cost for many-to-many traffic. | RiF in-flight tableMaintains millions of concurrent operations. |
| Ethernet adaptation layerBrings standard Ethernet to CN6000 silicon. | Ultra Ethernet compliant (Preview)Reflects the Cornelis commitment to open standards. | Priority Flow Control8 priorities for lossless traffic classes. |
| IEEE 1588 PTPPrecise time sync across the cluster. | SR-IOVShare one SuperNIC across virtual functions. | PCIe 6.0 x16Host bandwidth to match 800 Gb/s. |
| NVMe-oF over RoCEv2 (roadmap)Planned shared-flash access for data ingest. | Air or conduction coolingFit the server design you run. |
| GPUDirectDirect RDMA placement for NVIDIA and AMD accelerators. |
| Post-quantum Root of TrustSPDM 1.4 with ML-DSA-87 and ECDSA P-384. |
| Standard telemetryPLDM, MCTP and Redfish for system integration and firmware updates. |
| Integrated management processorMCTP over PCIe and SMBus; PLDM monitoring, firmware update and Redfish Device Enablement; SPDM 1.4. |
Specifications
The CN6000 Ethernet SuperNIC is a low-profile PCIe 6.0 x16 card with dual QSFP-DD ports, two 400 Gb/s data paths and RoCEv2 in hardware. It comes in air-cooled and conduction-cooled versions to fit the servers you run. See the product brief for full details.
| Protocols | RoCEv2; Ultra Ethernet compliant (Preview) |
|---|---|
| Bandwidth | 800 Gb/s each way; 1.6 Tb/s bidirectional |
| Data paths | Two 400 Gb/s |
| Ports | Dual x4-lane QSFP-DD, 112 Gb/s per lane |
| Host interface | PCIe 6.0 x16 |
| Queue pairs | Lightweight QPs; RiF table maintains millions of concurrent operations |
| Flow control | Priority Flow Control, 8 priorities (lossless) |
| Timing | IEEE 1588 PTP |
| Switches | Standards-based 400G and 800G Ethernet switches |
| Accelerators | GPUDirect for NVIDIA and AMD |
| Storage | NVMe-oF over RoCEv2 (roadmap) |
|---|---|
| Operating systems | Red Hat Enterprise Linux (RHEL); other distributions on request |
| Security | PQC Root of Trust; SPDM 1.4 |
| Management | SR-IOV; MCTP, PLDM, Redfish (RDE) |
| Form factor | PCIe low profile, 3.12 x 6.6 in (7.92 x 16.76 cm) |
| Brackets | Low profile installed; standard in package |
| Cooling | Air-cooled or conduction-cooled |
| Power (typ/max) | AOC 57.5/62.2 W; DAC 42.4/46.0 W |
| Weight | Air-cooled 0.62 lb (0.280 kg); conduction-cooled 0.52 lb (0.235 kg) |
Ordering information
CN6000 is sold through server OEMs and channel partners in configured systems. Ask Cornelis or a partner for current availability.
| Item Name | Ethernet SKU, air |
| Item number | Item no. |
| Description | CN6000 Ethernet SuperNIC, QSFP-DD, x16 PCIe 6.0, air-cooled, low profile, single |
| Weight | 0.62 lb (0.280 kg) |
| Item Name | Ethernet SKU, conduction |
| Item number | Item no. |
| Description | CN6000 Ethernet SuperNIC, QSFP-DD, x16 PCIe 6.0, conduction-cooled, low profile, single |
| Weight | 0.52 lb (0.235 kg) |
Safety, environment and operating conditions
Use these listings to confirm the SuperNIC fits your site conditions and your region's regulatory requirements before you plan a deployment.
| Temperature0 to 55 °COperating (derate 1 °C per 175 m above 900 m); storage -40 to 70 °C | Humidity10% to 85%Operating; storage 5% to 95%; non-condensing | Altitude0 to 3,200 mOperating; non-operating 0 to 12,000 m |
| US and Canada | NRTL 62368-1; CSA 22.2 No. 62368-1 |
| Europe | EN62368-1 |
| International | CB Scheme: IEC 62368-1 |
| RoHS | RoHS II Directive 2011/65/EU; Directive 2015/863 |
| REACH | REACH |
| US | FCC Part 15, Subpart B, Class A | Australia and New Zealand | AS/NZS CISPR 32, Class A |
| Canada | CAN ICES-3(A)/NMB-3(A) Issue 7 | Korea | KS C 9832 Class A; KS C 9835 |
| Europe | EN55032 (CISPR32), EN55035, EN61000-3-2, EN61000-3-3 | Taiwan | Taiwan |
| Japan | VCCI, Class A |
Frequently asked questions
-
What is the CN6000 Omni-Path SuperNIC?
It is an 800G network adapter for AI and HPC that runs the Omni-Path fabric. It is designed to deliver 1.6 billion messages per second and sub-microsecond MPI latency on a lossless fabric. -
What is the CN6000 Omni-Path Switch?
It is a 1U, liquid-cooled switch with 48 ports at 800G, or 96 at 400G, and 76.8 Tb/s of full-duplex bandwidth. It is natively designed for multi-plane fabrics: two independent 400G planes let one cable per node reach both. -
Is there a CN6000 SuperNIC for Ethernet networks?
Yes. Cornelis also offers the CN6000 Ethernet SuperNIC, which runs RoCEv2 on standards-based Ethernet switches and is Ultra Ethernet compliant (Preview). The CN6000 Omni-Path Switch is Omni-Path only. -
How does CN6000 Omni-Path help disaggregated inference?
Disaggregated inference splits prefill from decode, and attention from FFN or expert layers, onto separate nodes. KV cache then moves from prefill to decode for each request, and small activations move between attention and FFN layers for each token. CN6000 Omni-Path is designed to bring sub-microsecond latency, 1.6 billion messages per second and 800 Gb/s per node to that traffic. Credit-based flow control keeps the fabric lossless, so no transfer waits on a retransmit. -
Does it work with my existing CN5000 fabric?
Yes. It is backward compatible with CN5000 infrastructure and uses the same OPX provider and host driver, so CN5000 and CN6000 SuperNICs can share one Omni-Path fabric. Applications carry forward to CN7000 with a fully compatible software stack. -
Which switches does it connect to?
Connect it to the CN6000 Omni-Path Switch for a full 800G dual-plane fabric. It also works with existing CN5000 edge and director-class switches. -
Which software runs on CN6000 Omni-Path?
Cornelis OPX, an open-source stack built on libfabric, runs the SuperNIC and the switch. MPI, OpenSHMEM, NCCL, RCCL, UCX and PyTorch run on it, and multi-rail drives both data paths for up to 800 Gb/s per node.