Cornelis Claims 1.6 Billion Messages a Second With New CN6000 800G Line

Cornelis says its new CN6000 networking family can handle up to 1.6 billion messages per second on its Omni-Path SuperNIC. The company announced the 800-gigabit product line on November 12, positioning it as a way to keep expensive AI accelerators and supercomputers from sitting idle while they wait for data.

The claim comes with fine print. Cornelis's Omni-Path product brief specifies the 1.6-billion-message rate as bidirectional and aggregate. It is not a promise that a single application completes that many sequential requests every second.

The launch covers three pieces of hardware: Omni-Path and Ethernet adapters, plus an Omni-Path switch. Cornelis says hardware is available now for evaluation and early deployments, with partner systems expected in the first quarter of 2027.

What does 800G actually buy?

The 800-gigabit figure means 800 gigabits per second of data-carrying capacity. Divided by eight, that is a nominal 100 gigabytes per second before protocol overhead and other practical limits. It describes how much data can move, not how fast a chatbot generates words.

Network performance splits into three separate measures, and each addresses a different bottleneck:

• Bandwidth — how much data moves each second. It matters for moving large blocks of data.
• Message rate — how many separate messages the adapter can process each second. It matters for many small exchanges.
• Latency — how long one communication takes. It matters when the next calculation must wait for a message.

The message-rate claim targets the second bottleneck. Sending many tiny pieces of information forces the adapter to process many separate operations, even when those operations never fill the link's bandwidth.

How big is the latency win over InfiniBand?

In Cornelis's internal 8-byte Intel MPI ping-pong test, CN6000 reported 0.95 microseconds of latency versus 1.2 microseconds for NVIDIA InfiniBand XDR at 800G — roughly 21% lower. MPI, or Message Passing Interface, is the standard method parallel scientific programs use to communicate.

That test measured one particular small-message exchange. It does not establish that a training run, simulation, or inference service will finish 21% faster. Real outcomes depend on how often the application communicates, the sizes of its transfers, congestion, and the rest of the system.

What's in the platform?

A NIC connects a server to the network; "SuperNIC" is Cornelis's term for its specialized high-performance adapters. A switch connects server links and forwards data between them. The full arrangement of adapters, switches, links, and communication software is what buyers call a fabric.

CN6000 offers two fabric choices. Omni-Path is Cornelis's purpose-built interconnect for parallel computing. Ethernet targets customers who want high-performance communication while keeping compatible Ethernet switching infrastructure. InfiniBand, the technology in the benchmark comparison, is another specialized fabric, and its switches are not interchangeable with Omni-Path switches.

The distinction matters at purchase time. Installing a capable adapter does not automatically upgrade every switch, cable, or software path its traffic will hit.

How does the Ethernet adapter cut processor overhead?

The Ethernet SuperNIC implements RoCEv2 — Remote Direct Memory Access over Converged Ethernet, version two. RDMA lets network hardware move data between authorized memory regions on different machines without the host CPU copying every payload through the conventional networking path.

The CPU still establishes and manages communication; the benefit is pulling it out of the data movement itself. NVIDIA's RoCE documentation explains that version two carries RDMA traffic in IP and UDP headers, letting it cross routed IP networks. Efficient deployment still depends on network configuration and congestion behavior, not just the presence of an Ethernet port.

Cornelis's Ethernet product brief lists hardware offload, support for large numbers of simultaneous operations, and Priority Flow Control, which can pause selected traffic classes to avoid packet loss. The announcement puts Ultra Ethernet support on a software-update path — buyers should distinguish that planned capability from the initial RoCEv2 shipping feature set.

Why does inference care about network speed?

AI inference — using a trained model to produce an answer — includes prefill, which processes the prompt, and decode, which generates the response one token at a time. Some serving systems run the two stages on separate processor groups sized for their jobs. That approach, called disaggregated inference, creates a handoff: information computed while reading the prompt must reach the processors generating the answer.

That information includes the key-value cache, or KV cache, which stores intermediate attention data so the model can reuse previous work. It is neither the model's full set of weights nor a saved text transcript. NVIDIA's Dynamo documentation describes the sequence: process the prompt, transfer the KV cache, then generate tokens on the decode worker. Faster transfers make the separation more practical, though moving the cache adds work of its own — disaggregation pays off only when scheduling and hardware benefits outweigh the handoff cost.

What happens when links get busy?

On Omni-Path, CN6000 uses credit-based flow control: senders transmit according to the receiving side's available capacity. Adaptive routing spreads traffic over available paths, and multiple network planes provide separate routes through the fabric.

The Omni-Path brief describes two independent 400G data paths that together deliver the 800G aggregate. If one plane needs service, the other stays available — but losing part of the network does not preserve full original capacity.

The platform also ships a hardware root of trust, secure boot, and device attestation to verify firmware and device identity. Its use of post-quantum digital signatures covers authentication and integrity, not a blanket claim that every network payload is encrypted.

Should buyers chase the peak numbers?

CN6000 builds on Cornelis's earlier scale-out networking work, including its expanded AMD collaboration for scientific computing. The pitch is high-performance interconnects paired with a wider choice of processors and network environments.

For buyers, the comparison that counts is the time and cost to complete their own workload. A simulation that repeatedly exchanges small updates may benefit from lower latency and higher message rates; a service moving large caches may care more about sustained transfer speed and congestion handling. Both need an appropriate fabric, working software support, and qualified server configurations.

CN6000's peak figures identify capabilities worth testing on real workloads. The practical win, as the announcement itself frames it, is keeping more compute productive while actual applications run.