The high-performance networking market has long been dominated by two primary architectures: Ethernet, originally designed for general-purpose networking more than 50 years ago, and InfiniBand, developed 25 years ago for HPC environments. Both face fundamental limitations when applied to today’s AI and HPC workloads, where massive parallel processing demands often leave expensive compute resources underutilized due to network bottlenecks.
It’s a challenge that Intel recognized over a decade ago and attempted to solve with its Omni-Path architecture. In 2020, the core team behind Omni-Path at Intel spun out Cornelis Networks as a new vendor tasked with advancing the Omni-Path technology. Since spinning out from Intel, Cornelis has maintained business continuity by continuing to sell and support Intel’s legacy 100Gbps Omni-Path products.
Today, Cornelis is out with its first major platform since leaving Intel with the debut of the company’s CN5000 platform, providing a 400Gbps networking platform that aims to compete with Ethernet and InfiniBand.
“If you look at current AI clusters or an HPC cluster, you see compute utilization that sits below 30% in some cases, and … in the best architectures and best cases, it goes to 50% with lots of help, with bespoke, custom network engineering,” Lisa Spelman, CEO of Cornelis Networks, told Network World. “We consider ourselves very mission driven to better utilize all the compute in the world.”
srcset=”https://b2b-contenthub.com/wp-content/uploads/2025/06/Lisa-Spelman-Cornelis-Networks.jpg?quality=50&strip=all 4979w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Lisa-Spelman-Cornelis-Networks.jpg?resize=300%2C240&quality=50&strip=all 300w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Lisa-Spelman-Cornelis-Networks.jpg?resize=768%2C614&quality=50&strip=all 768w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Lisa-Spelman-Cornelis-Networks.jpg?resize=1024%2C819&quality=50&strip=all 1024w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Lisa-Spelman-Cornelis-Networks.jpg?resize=1536%2C1229&quality=50&strip=all 1536w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Lisa-Spelman-Cornelis-Networks.jpg?resize=2048%2C1638&quality=50&strip=all 2048w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Lisa-Spelman-Cornelis-Networks.jpg?resize=871%2C697&quality=50&strip=all 871w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Lisa-Spelman-Cornelis-Networks.jpg?resize=210%2C168&quality=50&strip=all 210w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Lisa-Spelman-Cornelis-Networks.jpg?resize=105%2C84&quality=50&strip=all 105w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Lisa-Spelman-Cornelis-Networks.jpg?resize=600%2C480&quality=50&strip=all 600w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Lisa-Spelman-Cornelis-Networks.jpg?resize=450%2C360&quality=50&strip=all 450w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Lisa-Spelman-Cornelis-Networks.jpg?resize=313%2C250&quality=50&strip=all 313w” width=”1024″ height=”819″ sizes=”(max-width: 1024px) 100vw, 1024px”>Lisa Spelman, CEO of Cornelis Networks
Cornelis Networks
Omni-Path architecture: beyond Ethernet and InfiniBand
The CN5000 represents a third architectural approach to high-performance networking, distinct from both Ethernet and InfiniBand implementations. Rather than attempting to retrofit existing protocols for AI and HPC workloads, Cornelis expanded Intel’s Omni-Path with these specific use cases in mind.
“If you think about the choices that the team could make, either at Intel or as a standalone company, it was work on changing Ethernet into something it wasn’t built to be, or build something from the ground up,” Spelman explained during a technical briefing. “What we did was get the architecture correct by [designing] for the workloads.”
The architecture incorporates several key technical differentiators designed specifically for scale-out parallel computing environments. Credit-based flow control ensures lossless data transmission, while dynamic fine-grained adaptive routing optimizes path selection in real-time. Enhanced congestion control mechanisms are designed to maintain consistent performance under heavy loads, which is a critical requirement for AI training workloads that can involve thousands of endpoints.
Performance metrics and benchmarking
Cornelis positions the CN5000’s advantages in specific technical metrics that address known bottlenecks in AI and HPC workloads. The company claims 2X higher message rates and 35% lower latency compared to other 400Gbps solutions.
What’s different about the Cornelis architecture is that with the same bandwidth, you can achieve double the message rates, Spelman explained. “To me, that’s the way that the architectural correctness for the workloads shows up.”
For AI workloads specifically, the company highlights 6X faster collective communication performance compared to remote direct memory access (RDMA) over converged Ethernet (RoCE) implementations. Collective operations like all-reduce functions represent critical bottlenecks in distributed training, where thousands of nodes must synchronize gradient updates efficiently.
The architecture’s congestion management becomes particularly relevant in AI training scenarios, where synchronized communication patterns can overwhelm traditional networking approaches. Omni-Path’s credit-based flow control and adaptive routing aim to maintain consistent performance even under these demanding conditions.
“With the exact same compute installed and just a swap of the network from another 400 gig to CN5000, you see application performance that improves by 30%,” Spelman said. “Normally to improve by 30% on an application’s performance, you would need a new CPU generation.”
srcset=”https://b2b-contenthub.com/wp-content/uploads/2025/06/Cornelis_Director-Class-Switch_Air-Cooled.jpg?quality=50&strip=all 5889w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Cornelis_Director-Class-Switch_Air-Cooled.jpg?resize=240%2C300&quality=50&strip=all 240w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Cornelis_Director-Class-Switch_Air-Cooled.jpg?resize=768%2C960&quality=50&strip=all 768w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Cornelis_Director-Class-Switch_Air-Cooled.jpg?resize=819%2C1024&quality=50&strip=all 819w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Cornelis_Director-Class-Switch_Air-Cooled.jpg?resize=1229%2C1536&quality=50&strip=all 1229w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Cornelis_Director-Class-Switch_Air-Cooled.jpg?resize=1638%2C2048&quality=50&strip=all 1638w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Cornelis_Director-Class-Switch_Air-Cooled.jpg?resize=558%2C697&quality=50&strip=all 558w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Cornelis_Director-Class-Switch_Air-Cooled.jpg?resize=134%2C168&quality=50&strip=all 134w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Cornelis_Director-Class-Switch_Air-Cooled.jpg?resize=67%2C84&quality=50&strip=all 67w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Cornelis_Director-Class-Switch_Air-Cooled.jpg?resize=384%2C480&quality=50&strip=all 384w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Cornelis_Director-Class-Switch_Air-Cooled.jpg?resize=288%2C360&quality=50&strip=all 288w, https://b2b-contenthub.com/wp-content/uploads/2025/06/Cornelis_Director-Class-Switch_Air-Cooled.jpg?resize=200%2C250&quality=50&strip=all 200w” width=”819″ height=”1024″ sizes=”(max-width: 819px) 100vw, 819px”>
Cornelis Networks
Current implementation and future interoperability strategy
The CN5000 currently implements Omni-Path as a complete end-to-end solution, requiring both Cornelis SuperNICs and switches to achieve full performance benefits. However, this represents just the first phase of a broader strategy designed to bridge proprietary performance with industry-standard interoperability.
“The CN5000 is an end-to-end network that the Super NIC and the switch, or the director class switch, work together,” Spelman explained. “It won’t be used in a split apart fashion.”
The CN5000 platform supports deployments scaling up to 500,000 endpoints, positioning it for the massive installations typical of national laboratories and enterprise AI initiatives. The hardware portfolio includes both single and dual-port SuperNICs with air and liquid cooling options, 48-port switches for smaller deployments, and 576-port director-class systems designed for spine-and-leaf topologies in larger installations.
The software stack demonstrates how Cornelis balances proprietary optimization with open standards. At the physical layer, Omni-Path implements its credit-based flow control and adaptive routing. Above this sits the OpenFabrics Alliance software layer, an open-source abstraction that Cornelis helped create and continues to support.
“We’ve used that open-source software layer as a way to make it incredibly easy for any customer who has used InfiniBand before or has used Ethernet before, to get started to deploy and move forward on the Omni-Path architecture,” Spelman said.
This open-source middleware handles the translation between applications and the underlying Omni-Path protocols, allowing existing HPC and AI software to run without modification. The OpenFabrics layer has been adopted by the Ultra Ethernet Consortium as a foundation component, demonstrating its industry acceptance beyond Cornelis deployments.
Looking ahead, future product generations will expand this interoperability approach.
“As we add Ethernet into the SuperNIC, you will then have [the option to use] the Cornelis switch, or you will be able to use another Ethernet switch,” Spelman noted.
In terms of next steps, the CN6000 (800Gbps) will incorporate dual-mode capabilities, supporting both native Omni-Path for maximum performance and Ethernet protocols for broader ecosystem compatibility. After that the CN7000 (1.6Tbps) will further integrate Ultra Ethernet Consortium standards while maintaining Omni-Path’s architectural advantages.
Spelman noted that Cornelis Networks’ first customer deployments are at the Texas Advanced Computing Center and with the U.S. Department of Energy.
“We look forward to scaling across global industries in the areas of automotive, energy discovery, oil and gas, health and life sciences, and any enterprise driving their initial and growing efforts in AI,” she said.
Source:: Network World