Cornelis Networks, a recognized leader in high-performance networking solutions, has revealed new details about its CN6000 SuperNIC, a next-generation 800 Gbps networking adapter designed to handle the demanding workloads of artificial intelligence (AI) and high-performance computing (HPC). With the ability to process up to 1.6 billion messages per second, the CN6000 represents Cornelis’ first effort to extend its Omni-Path architecture to Ethernet, including both RoCEv2 and Ultra Ethernet standards. The adapter is engineered to accelerate GPU utilization, reduce data center energy consumption, and deliver faster results for mission-critical computing tasks.

Lisa Spelman, CEO of Cornelis, highlighted the significance of the launch, stating that as AI infrastructures grow to massive scales, traditional Ethernet solutions encounter critical bottlenecks. The CN6000 addresses this by combining proven high-performance architecture with unlimited RoCEv2 scalability and Ultra Ethernet support, providing a single platform capable of sustaining performance at unprecedented levels.

Meeting the Demands of Large-Scale AI

The CN6000 is part of Cornelis’ broader strategy to offer 800 Gbps networking solutions tailored for organizations operating at scale. Compared to its predecessor, the CN5000, the CN6000 doubles performance and incorporates innovations that make a tangible difference for AI and HPC workloads.

A key feature is the Multi-Modal Performance Pipeline, which enables the CN6000 to achieve its full 800 Gbps throughput while maintaining extremely low latency and an industry-leading message rate. This combination allows data centers to maximize GPU utilization, accelerate AI training, and reduce both power consumption and operational costs.

Breaking Through RoCEv2 Limits

Traditional RoCEv2 networks often struggle in large GPU clusters due to the memory demands of managing queue pairs, leading to unpredictable latency and reduced throughput in complex communication patterns. Cornelis overcomes these limitations with a lightweight queue pair design and a hardware-accelerated RoCEv2 In-Flight (RiF) table, capable of tracking millions of concurrent operations with minimal memory overhead. This ensures predictable latency and consistent performance, even in massive deployments.

Supporting Next-Generation Ultra Ethernet

The CN6000 is fully compatible with Ultra Ethernet, a standards-based approach developed by leading tech companies for AI-scale networking. Cornelis’ involvement in creating and contributing to libfabric, the core interface adopted by Ultra Ethernet applications, ensures seamless software integration and long-term compatibility.

Robert Hormuth, Corporate VP at AMD, emphasized the importance of open, standards-driven innovation for scalable AI and HPC infrastructures. He noted that the Ultra Ethernet ecosystem allows customers to maintain flexibility while meeting evolving compute requirements.

Industry Partnerships Reinforce Performance

Cornelis is aligning the CN6000 with leading computing platforms to ensure smooth integration. Collaborations with Intel focus on pairing CN6000 capabilities with the Xeon platform, optimizing RDMA scalability and low-latency performance for AI and HPC workloads. Lenovo, Altair, Synopsys, Atipa Technologies, Nor-Tech, Microway, PSSC Labs, and SourceCode all highlighted the CN6000’s potential to deliver high performance and predictable, congestion-free operations in demanding environments. These endorsements underline the CN6000’s versatility and scalability for modern AI and simulation tasks.

Expanding the CN6000 Ecosystem

Beyond the SuperNIC itself, Cornelis will introduce CN6000-series Omni-Path switches and director-class platforms, forming a complete end-to-end solution for HPC and AI networks. Customer sampling is slated for mid-2026, with production deployments to follow, positioning the CN6000 as the cornerstone of next-generation 800 Gbps infrastructure.

About Cornelis Networks

Cornelis Networks develops networking solutions designed to accelerate AI and HPC workloads. Its Omni-Path architecture enables lossless, congestion-free communication, improving compute efficiency and reducing AI training times. Cornelis’ solutions are deployed across commercial, academic, government, and cloud environments, powering complex tasks ranging from foundation model training to climate modeling and real-time analytics.

For more information about Cornelis and the CN6000 SuperNIC, visit Cornelis Networks.