In the previous batch we covered CN5000: Cornelis and AMD announced their joint HPC program on 2026-11-11, anchored by GWDG's EMMY P4 in Göttingen — about 300 dual-socket EPYC nodes running on CN5000 Omni-Path — and the next generation teased then has now formally landed. Cornelis announced the CN6000 family on 2026-11-12 (Wayne, Pa.), doubling Omni-Path bandwidth to 800G; StorageReview reported on 2026-11-16, and the family is on display at SC26 (2026-11-16 to 2026-11-19, Chicago). CEO Lisa Spelman's opening line is still that same bill: "Every minute an accelerator waits on the network is capacity our customers paid for and didn't use."
The trio: SuperNIC, switch, Ethernet SuperNIC

The CN6000 family has three parts: an Omni-Path SuperNIC, a 48-port Omni-Path switch, and an Ethernet SuperNIC that runs RoCEv2 against standard 800G Ethernet switches. The adapters, switch and software are available now for evaluation and early deployments, with systems from OEM and system partners expected from 2027 Q1. Key specifications (vendor figures):
| Component | Key specifications |
|---|---|
| Omni-Path SuperNIC | Low-profile PCIe 6.0 x16; 2 QSFP-DD ports at 400Gb/s each, 800Gb/s total; 1.6B messages/s (design target, bidirectional); sub-microsecond MPI latency (design target); typical power 42.4W (passive copper) / 57.5W (active optical) |
| Omni-Path switch | 1U; 48 QSFP-DD ports at 800G, or 96 ports at 400G; 76.8Tb/s full duplex; can run as 2 independent 48-port 400G planes; switch latency as low as 105ns; typical power 856W (all copper) / 1663W (all optical); 240-277VAC or 48VDC, N+N hot-swap power supplies |
| Ethernet SuperNIC | Same low-profile PCIe 6.0 x16, dual QSFP-DD; RoCEv2 to standard Ethernet switches; PFC across 8 priorities, IEEE 1588 PTP, SR-IOV, GPUDirect (NVIDIA/AMD); Ultra Ethernet in preview (via software updates); NVMe-oF over RoCEv2 on the roadmap (no date); RHEL support |
On the Omni-Path side, multi-plane/multi-rail configurations isolate traffic and keep the fabric running through link failures, with credit-based flow control for lossless delivery; a hardware Root of Trust verifies firmware before the adapter joins the system, using SPDM 1.4 attestation and ML-DSA post-quantum signatures; Cornelis puts Omni-Path fabric scale at more than 500,000 endpoints. The software stack is identical to CN5000 (NCCL, RCCL, libfabric, MPI, OpenSHMEM, same host driver), so existing sites adding 800G hardware keep their software base. Paired with the SuperNIC, the switch enables adaptive routing to spread traffic across available paths and ease congestion.
A path per core: spreading the message rate across cores

The claim from the Cornelis-AMD joint program: CN6000 can give every core in a 512-core dual-socket server a dedicated network connection (vendor claim). Spread 1.6B messages/s across 512 cores and you get about 3.1M messages/s per core (calculated) — for fine-grained MPI workloads (neighbor communication in CFD, weather, life sciences), that "per-core budget" explains application scaling better than single-port bandwidth.
The host-interface math is worth doing too: 800Gb/s equals 100GB/s (per direction); PCIe 5.0 x16 theoretically delivers about 64GB/s — not enough; PCIe 6.0 x16 at about 128GB/s leaves headroom (theoretical calculation). That is why CN6000 must move to PCIe 6.0 — before the network doubles, the host bus doubles first. Likewise, 48×800G full duplex on the switch gives 76.8Tb/s — internally consistent.
Versus XDR InfiniBand: how to read the 21% latency edge

Cornelis says that at the same 800G speed, CN6000's MPI latency is 21% lower than NVIDIA InfiniBand XDR: internal testing with Intel MPI ran an 8-byte ping-pong, where the CN6000 Omni-Path SuperNIC measured 0.95µs against 1.2µs for XDR. These are vendor results; StorageReview did not run or audit them, and the test conditions behind the message rate are unpublished. One calibration update: the "23% lower" figure our previous CN5000 article cited was a projection at the time; this launch's measured claim is 21%.
Read it soberly: an 8-byte ping-pong is the most favorable scenario for latency marketing, and real application gains depend on message-size distribution, hop count and congestion; the 0.25µs gap per exchange accumulates across many hops at scale, but the comparison topology is unpublished and cannot be extrapolated directly. Ecosystem and ops-tooling maturity remain InfiniBand's moat; however, CN6000 lists NCCL/RCCL support, a substantive win for AI training clusters — this is no longer a purely HPC route war.
Copper versus optics: a power bill that nearly doubles
Cabling choice is power choice. The switch draws typically 856W with all-copper cabling and 1663W with all active optical cables, about 1.9 times (calculated); the SuperNIC follows the same pattern, 42.4W versus 57.5W. DAC copper inside the rack, optics between racks, is the standard power-saving play in HPC machine rooms; the 240-277VAC or 48VDC supply options and N+N hot-swap power supplies also fit the distribution reality of high-density rooms.
The 2027 Q1 OEM window, and procurement lessons for research/HPC clusters
The platform validation net is wide: CN6000 supports AMD, Arm-based and Intel CPUs with AMD, NVIDIA and Qualcomm accelerators, is being validated with AMD EPYC 9006 and next-generation Instinct platforms, and the Intel collaboration extends to CN6000 and next-generation Xeon. The OEM/system partner list includes Dell, Hyve Solutions, Lenovo, NEC, Penguin Solutions and Supermicro: Lenovo has qualified CN5000 across ThinkSystem V3/V4 and is qualifying Cornelis hardware for its new AMD EPYC ThinkSystems; Supermicro has validated CN5000 on FlexTwin, BigTwin and Hyper and plans to extend to CN6000; Penguin Solutions plans to validate and support CN6000 in its AI/HPC systems. Cornelis calls this its second consecutive "first-silicon-to-production" generation (vendor claim).
For Pearl River Delta universities, research institutes and HPC cluster buyers, we extend the previous batch's advice with three additions. First, evaluation units are available now but complete systems wait until 2027 Q1 — write "network and compute accepted together" into the tender and run a POC on your own MPI workloads; don't judge by 8-byte latency alone. Second, the same stack and driver as CN5000 means older clusters can mix generations smoothly, and multi-OEM supply reduces single-source risk. Third, ML-DSA post-quantum signatures and the hardware Root of Trust are genuine pluses for long-lifecycle, sensitive research clusters — worth writing into security requirements. The route choice (InfiniBand/Omni-Path/RoCE) ultimately depends on your codes, topology and ops team, not a single benchmark.
Sources & verification notes (accessed 2026-11-22)
- StorageReview (Harold Fritts, 2026-11-16): https://www.storagereview.com/news/cornelis-cn6000-800g-omni-path-supernic-switch (fetched and verified directly)
- Cornelis official press release (2026-11-12): https://www.cornelis.com/stories/cornelis-cn6000-networking-platform-delivers-industry-leading-1-6-billion-messages-per-second-at-800g-for-ai-and-hpc (JS-rendered page; headline/date and key content verified via search-engine snapshots; the Business Wire syndication timed out, full text not retrieved)
- Cornelis CN6000 product page: https://www.cornelis.com/products/cn6000 (JS-rendered page; only the title verifiable)
- Our previous CN5000 article (2026-11-15 batch, covering the 2026-11-11 AMD joint program and the GWDG case) is the series predecessor.
- Calculated: 1.6B÷512≈3.1M messages/s per core, 800Gb/s=100GB/s, PCIe 5.0 x16≈64GB/s, PCIe 6.0 x16≈128GB/s, 1663W/856W≈1.9x, the 0.25µs gap. Vendor claims/unaudited: 1.6B messages/s and sub-microsecond latency (design targets), 21% (0.95µs vs 1.2µs, 8-byte Intel MPI internal test), the per-core path on 512 cores, >500,000 endpoints, 105ns, 42.4W/57.5W, 856W/1663W, second consecutive first-silicon-to-production. 23% is the older projection cited in the previous batch. OEM systems in 2027 Q1 are a vendor plan.
