Cornelis Networks Unveils OmniPath Hardware to Connect AI Chips at Scale
Cornelis Networks Unveils New Networking Hardware to Connect AI Chips
Cornelis Networks introduced a new suite of networking hardware and software designed to link up to half a million artificial intelligence (AI) chips, as reported by Reuters. This release aims to address a critical challenge in AI datacenters, ensuring fast and reliable data transfer between chips in large-scale AI computations.
Solving the AI Datacenter Networking Bottleneck
AI computing chips are known for their impressive speed, but as AI datacenters scale, the network links between these chips struggle to keep up with the data demands.
Cornelis Networks addresses this problem by offering a new network technology called OmniPath. Unlike traditional network technologies like Ethernet or the InfiniBand protocol used by Nvidia through its acquisition of Mellanox, OmniPath is tailored specifically to the high-performance needs of AI computing.
The company's new CN5000 networking chips will provide the backbone for AI datacenters, ensuring faster and more efficient data flows between chips, allowing for large-scale AI operations to function more smoothly and at higher performance levels.
Partnerships and Customers
The CN5000 chips, powered by OmniPath technology, will initially be shipped to clients like the US Department of Energy starting in the third quarter of 2025. Although Cornelis has the support of Intel, its network chips are made to be compatible with AI chips from a variety of different companies, such as Nvidia and AMD.
According to Reuters, Cornelis plans to release the next generation of its chips in 2026, which will offer compatibility with Ethernet networks. This move is expected to alleviate concerns from customers worried about being locked into proprietary technology, providing them with a more flexible and scalable solution.
Also read: Snowflake Summit 2025: New AI Tools SnowConvert & Cortex AISQL Unveiled
A New Approach to AI Datacenter Networking
Cornelis CEO Lisa Spelman emphasized that the company's technology aims to break free from the limitations of 45-year-old architecture and 25-year-old networking solutions that have been traditionally used in AI and high-performance computing. The company seeks to provide customers with better economic performance without compromising AI chip performance.
