Skip to the stories
Windows & PowerShell

Cornelis CN6000 Pushes Omni-Path to 800G With a New SuperNIC and 48-Port Switch

Cornelis CN6000 doubles Omni-Path to 800G with a SuperNIC, a 48-port switch and an RoCEv2 Ethernet SuperNIC. Partner systems are expected in Q1 2027.

Cornelis CN6000 Pushes Omni-Path to 800G With a New SuperNIC and 48-Port Switch

Cornelis has put a number on the cost of a slow network: idle accelerators. CEO Lisa Spelman framed the CN6000 launch around the idea that every minute a GPU waits on the fabric is "capacity our customers paid for and didn't use." The CN6000 family doubles the bandwidth of the company's CN5000 generation, and it arrives in three parts: an Omni-Path SuperNIC, a 48-port Omni-Path switch, and an Ethernet SuperNIC. For teams building AI and HPC clusters, the practical question is which of those parts belongs in the next design review.

Cornelis CN6000 SuperNIC, a low-profile PCIe card with a copper heatsink and a QSFP-DD port cage

What the Omni-Path SuperNIC Brings

The Omni-Path SuperNIC is a low-profile PCIe 6.0 x16 card with two QSFP-DD ports. Each port carries a 400Gb/s data path, which adds up to 800Gb/s per card. Cornelis rates the adapter at 1.6 billion messages per second, which it describes as the industry's highest, and at sub-microsecond MPI latency. Both numbers are design targets, and the company describes the message rate as bidirectional. It hasn't published the test conditions behind the message-rate figure, so treat that number as a starting point for your own benchmarking rather than a purchasing baseline.

Cornelis also claims MPI latency 21 percent lower than NVIDIA InfiniBand XDR at the same 800G speed. That comparison comes from internal Cornelis testing of an 8-byte MPI ping-pong run with Intel MPI. The CN6000 SuperNIC measured 0.95µs, against 1.2µs for InfiniBand XDR. Those are vendor results, and nobody outside Cornelis has reproduced them yet. Here's the thing: for latency-sensitive collectives, a gap of that size is worth checking on your own applications, not just on a ping-pong microbenchmark.

Power is where facilities teams should look first. Typical draw is 42.4W with passive copper cables and 57.5W with active optical cables. Multiply that across a node count and the NIC power becomes a line item in your rack and PDU planning, even though it's small next to the accelerators.

On the fabric side, the SuperNIC supports multi-plane and multi-rail configurations. Cornelis says these let operators isolate traffic and keep a fabric running through link failures. It uses credit-based flow control for lossless delivery. A dedicated hardware Root of Trust checks firmware before the adapter joins the system, using SPDM 1.4 attestation and ML-DSA post-quantum signatures. Cornelis puts the scale of an Omni-Path fabric at more than 500,000 endpoints.

The 48-Port Switch and Its Power Envelope

Cornelis CN6000 Omni-Path switch diagram showing two 48-port 400G planes, with each node connected to both planes for up to 800Gb/s

The CN6000 Omni-Path switch is a 1U box with 48 QSFP-DD ports running at 800G, or 96 ports at 400G. Full-duplex capacity is 76.8Tb/s. It can also operate as two independent 48-port 400G planes, which matches the dual-plane layout shown in Cornelis's diagram, where each node connects to both planes for up to 800Gb/s. Cornelis lists switch latency as low as 105ns. When the SuperNIC and switch are paired, adaptive routing spreads traffic across available paths to reduce congestion.

The switch draws 856W with all-copper cabling and 1,663W with active optical cables on every port. It accepts 240-277VAC or 48VDC through hot-swap N+N power supplies. That 1,663W figure with optics fully populated is a number to carry into a capacity review before anyone orders switch hardware. Cabling choice moves the power budget a lot, so decide between copper and optics early.

Ethernet Option and Software Stack

The Ethernet SuperNIC uses the same low-profile PCIe 6.0 x16 form factor and dual QSFP-DD ports. It connects to standard Ethernet switches over RoCEv2, which matters for sites already running 800G Ethernet fabrics and not planning a second network. Cornelis lists Priority Flow Control across eight priorities, IEEE 1588 PTP, SR-IOV, and GPUDirect support for NVIDIA and AMD accelerators.

Two features are still ahead. Ultra Ethernet support is in preview and will arrive through software updates. NVMe-oF over RoCEv2 is on the roadmap. Cornelis hasn't given dates for either, so don't build a project plan around them yet. For the Ethernet SuperNIC, Red Hat Enterprise Linux is the listed distribution, with others available on request.

On the Omni-Path side, CN6000 supports NCCL, RCCL, libfabric, MPI, and OpenSHMEM. It uses the same software stack and host driver as CN5000. For sites already running the earlier generation, that's the part that reduces friction. Moving to 800G hardware doesn't have to mean re-validating the entire software base.

Platform Support and Timing

Cornelis says CN6000 supports AMD, Arm-based, and Intel CPU platforms, along with AMD, NVIDIA, and Qualcomm accelerators. The company is validating the family with AMD EPYC 9006 Series CPUs and next-generation AMD Instinct GPU platforms. It is also extending its Intel collaboration to CN6000 and next-generation Xeon platforms.

OEM and system partners planning validation and support include Dell Technologies, Hyve Solutions, Lenovo, NEC, Penguin Solutions, and Supermicro. Lenovo has qualified CN5000 across its ThinkSystem V3 and V4 servers and is now qualifying Cornelis hardware with its new AMD EPYC-powered ThinkSystem servers. Supermicro has validated CN5000 on its FlexTwin, BigTwin, and Hyper platforms and plans to extend that work to CN6000. Penguin Solutions plans to validate and support CN6000 in its AI and HPC systems.

Cornelis says the adapters, switch, and software are available now for evaluation and early deployments. Systems from OEM and system partners are expected to begin in Q1 2027. That gap matters for procurement. If your cluster build is landing in the next two quarters, evaluation hardware is the realistic path, and production OEM configurations come later. Cornelis is demonstrating the family at SC26 in Chicago, which gives teams a chance to see the hardware before committing.

My read is that the CN6000 makes the most sense for sites already invested in Omni-Path, where the matching software stack lowers the cost of the upgrade. Teams weighing Ethernet against InfiniBand will want the Ethernet SuperNIC on their shortlist, but should wait for independent results before trusting the latency claims. The next useful data points are third-party benchmarks and the power numbers measured under real loads.

Cornelis CN6000

Source: Harold Fritts · www.storagereview.com

Same flag