Home

Cornelis Unveils CN6000 800 Gbps Ethernet SuperNIC for AI and HPC at Scale

via Business Wire

First SuperNIC to bring together Omni-Path, Ethernet RoCEv2, and Ultra Ethernet for maximum performance, flexibility and scalability

Cornelis, a leading provider of intelligent, high-performance networking solutions, is providing new details on the CN6000 SuperNIC, an 800 Gbps networking adapter purpose-built to meet the performance and scalability demands of AI and high performance computing. Delivering the industry’s highest performance with up to 1.6 billion messages per second, the CN6000 extends Cornelis’ leadership in high-performance interconnects to Ethernet RoCEv2 and Ultra Ethernet, bringing Cornelis’ proven Omni-Path architecture to the world’s most widely used networking standard. The CN6000 accelerates AI training and HPC workloads alike, enabling higher GPU utilization, lower data center power and cooling costs, and faster time-to-results for mission-critical computing.

“As AI factories scale toward gigawatt-sized deployments, traditional Ethernet RoCEv2 architectures are reaching their limits,” said Lisa Spelman, CEO, Cornelis Networks. “With the CN6000 SuperNIC, we’re eliminating that bottleneck by bringing our proven high-performance architecture to Ethernet for the first time — combining limitless RoCEv2 scalability with support for Ultra Ethernet and Omni-Path. One platform. Infinite scale. No compromises.”

“The industry is at an inflection point where AI clusters demand Ethernet compatibility without giving up the deterministic performance of purpose-built HPC fabrics,” said Sameh Boujelbene, Vice President at Dell’Oro Group. “By extending its Omni-Path innovations to an 800 Gbps Ethernet environment, Cornelis is addressing the performance, scalability, and interoperability gaps that have challenged operators building large-scale GPU fabrics.”

Built to Scale: An Architecture for Gigascale AI

The first solution in Cornelis’ broader 800 Gbps networking family, the CN6000 SuperNIC is purpose-built for organizations that need performance at scale without compromising flexibility or efficiency. The CN6000 delivers 2X the performance of the previous-generation CN5000 and introduces several key innovations that redefine AI and HPC networking:

  • Multi-Modal Performance Pipeline — The CN6000 delivers up to 1.6 billion messages per second and sustains 800 Gbps of bandwidth, combining ultra-low latency, the industry’s highest message rate, and maximum throughput in a single architecture. By unlocking full bandwidth and consistent low latency at scale, the CN6000 helps organizations accelerate AI training, maximize infrastructure utilization, and achieve faster results — all while reducing data center power and operational costs.
  • Limitless RoCEv2 Scalability — Traditional RoCEv2 architectures struggle to scale in large GPU clusters due to the memory-heavy overhead of managing queue pairs (QPs), leading to performance bottlenecks and unpredictable behavior in all-to-all communication patterns. The Cornelis CN6000 overcomes this limitation with a fundamentally different design: leveraging lightweight QPs and a hardware-accelerated RoCEv2 In-Flight (RiF) table capable of tracking millions of concurrent RoCEv2 operations with minimal footprint. This ensures predictable latency and full bandwidth at any scale—delivering truly limitless RoCEv2 without architectural compromises.
  • Ultra Ethernet — Fully compliant with Ultra Ethernet and RoCEv2, the CN6000 provides a standards-based path to next-generation 800 Gbps Ethernet networking. Cornelis is not just supporting the Ultra Ethernet software ecosystem — we helped build it. As the original developer and a key contributor to libfabric, the core software interface now being adopted for Ultra Ethernet AI applications, Cornelis ensures seamless integration and long-term compatibility.

"We believe open, standards-based innovation is key to unlocking scalable, high-performance infrastructure for the future of AI and HPC,” said Robert Hormuth Corporate Vice President, Architecture and Strategy, AMD. “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.”

The CN6000 is designed to complement next-generation compute platforms ensuring seamless integration for AI and HPC workloads at scale. In collaboration with Intel, Cornelis is working to align CN6000 innovations with the Intel Xeon platform to help customers achieve predictable low latency, massive RDMA scalability, and standards-based Ethernet performance for future AI and HPC deployments. "Cornelis’ CN6000 and our Xeon Platform share a common vision: delivering uncompromising performance for AI and HPC,” said Srini Krishna, Intel Fellow, Data Center Products. “Features like limitless RoCEv2 and Ultra Ethernet will play a critical role in scaling future workloads efficiently and reliably.”

"Lenovo is focused on building HPC and AI-optimized systems that scale,” said Scott Tease, VP and GM, HPC and AI, Lenovo. “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.”

"As engineering simulation evolves to meet the demands of digital twins, AI integration, and increasing model complexity, advanced interconnect solutions like the CN6000 will play a pivotal role in enabling Altair customers to achieve their most ambitious goals," said Eric Lequiniou, Vice President of Altair Radioss Development and Altair Solver HPC.

“Our collaboration with Cornelis Networks underscores our shared commitment to delivering scalable, high-performance solutions for engineering simulation,” said Wim Slagter, Senior Director, Partner Programs at Synopsys. “As simulation continues to evolve – driven by digital twins, AI integration, and growing model complexity – advanced interconnect technologies like the CN6000 will be essential to maintaining the compute and communication balance that empowers Synopsys customers to achieve their most ambitious engineering goals.”

“We greatly value our partnership with Cornelis,” said Bart Willems, CTO, Atipa Technologies. “Their networking solutions consistently deliver exceptional performance and scalability. The CN6000, now supporting 800G Ultra Ethernet, builds on a trusted platform as we prepare for our next phase of growth.”

“Our deployment of Cornelis products has grown exponentially this year,” said Bob Dreis, Senior HPC Account Manager, Nor-Tech. “Products like the new CN6000 platform consistently offer the performance and value our customers demand. We are excited about the CN6000 platform leading us into the 800G generation.”

“Congestion-free operation is crucial for fabrics addressing both AI and simulation, and the Cornelis CN6000 platform promises to provide a consistent, low-latency fabric necessary to keep these workloads running efficiently into the 800G fabric era,” said Brett Newman, VP Marketing, Microway, Inc. “Cornelis has delivered the performance and predictability required for many of our important AI and HPC customers.”

“As our compute footprint has grown, maintaining high utilization across our accelerators has become critical,” said Alex Lesser, Vice President, Sales and Marketing, PSSC Labs. “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.”

“Trusted for over 30 years and serving 40 countries, SourceCode has been a force behind many co-designed HPC/AI solutions, architecting, installing and servicing clusters using the latest technologies,” said Matt Ritter, Vice President of Engineering at SourceCode. “If the network is the computer, then the fabric is the supercomputer. Consistent low-latency and high bandwidth are the foundation of efficiency and price/performance. Today, SourceCode and Cornelis can offer a new solution that raises the bar. Built around the Cornelis CN6000, it enables scalable or already-large HPC and AI clusters to deliver more for less, running efficiently into the 800G infrastructure upgrade cycle."

In addition to the multi-protocol CN6000 SuperNIC, Cornelis will also offer a new family of CN6000-series Omni-Path switches and director-class platforms, enabling an end-to-end Omni-Path solution purpose-built for HPC and AI.

CN6000 customer sampling is expected to begin mid-2026, with production deployments to follow.

Visit us at SC25 at Booth #2009 to learn how we’re redefining performance, scalability, and interoperability for AI and HPC at scale. To learn more about Cornelis and its high-performance networking solutions please visit: https://www.cornelisnetworks.com/.

About Cornelis

Cornelis Networks delivers high-performance, scale-out networking solutions that accelerate AI and HPC workloads. Built on the powerful Omni-Path architecture, Cornelis technology enables lossless, congestion-free networking that reduces training time, improves inference, and maximizes compute utilization. From foundation model training to complex climate modeling and real-time analytics, Cornelis' solutions power the most demanding workloads across commercial, academic, government and cloud environments. With a focus on performance, scalability, and efficiency, Cornelis helps organizations achieve faster insights and greater return on infrastructure investments. Learn more at cornelisnetworks.com.

Altair is part of Siemens Digital Industries Software. To learn more, please visit altair.com or sw.siemens.com.

Frequently asked questions about Cornelis’ Omni-Path architecture and the pain points it addresses:

  1. What is Cornelis' Omni-Path architecture?

    Cornelis' Omni-Path architecture is a high-performance interconnect fabric designed for HPC and AI workloads. It delivers high bandwidth, high message rates, and low latency through enhanced hardware and open-source software. Its key features include lossless, congestion-free data flow via adaptive routing and credit-based flow control, and support for diverse HPC and AI applications across multiple hardware vendors.
  2. What are RoCEv2 and Ultra Ethernet environments and how do they differ?

    RoCEv2 is today's standard for high-speed data movement over Ethernet. It offers strong performance but requires careful tuning to maintain a lossless network. Ultra Ethernet, developed by leading tech companies through the Ultra Ethernet Consortium, is a next-generation approach designed for the AI era. It delivers similar high performance with less complexity, using smarter congestion management to scale efficiently across massive computing clusters. In short: RoCEv2 powers today's networks, while Ultra Ethernet is built for tomorrow's.
  3. What is RDMA/RoCEv2 and why does it struggle with AI workloads?

    RDMA (Remote Direct Memory Access) moves data directly between computers' memory without using the CPU, making transfers fast and efficient. It's used in technologies like RoCE (RDMA over Ethernet) and works well for predictable, sequential data transfers. However, AI workloads (especially large-scale deep learning) involve irregular, massive, and highly parallel communication between GPUs. Traditional RoCEv2 optimization for steady, predictable transfers makes it less effective at handling the unpredictable and heavy communication demands of modern AI training, creating potential bottlenecks in high-performance setups.
  4. How does Cornelis' Omni-Path architecture address AI and HPC bottlenecks?

    Cornelis' Omni-Path architecture solves AI and HPC bottlenecks through its lossless, congestion-free network fabric. It uses credit-based flow control and dynamic adaptive routing to prevent data flow interruptions, maximizing throughput and lowering latency. By actively managing congestion and allocating memory in advance, it delivers more predictable performance and eliminates the delays that cause traditional networks to degrade under heavy load. This is critical for data-intensive workloads like AI training and large-scale simulations.

Contacts