Cornelis Networks, which supplies networking equipment for AI data centers and scientific computing, announced approximately $205 million in financing on September 14. Alongside the funding, it introduced Active Compute Fabric, an architecture intended to help connected processors spend more time doing useful work.
The proposed workflow moves some calculations into the network itself. Cornelis marketing chief Brandon Draeger told SiliconANGLE that results from multiple processors can be combined as they travel, so the destination receives an assembled result. Training updates can also be compressed in transit. The intended benefit is less traffic and less time spent by AI chips communicating and synchronizing their work.
Cornelis is extending its business toward connections closer to the accelerators inside rack-scale AI systems. Its architecture uses open standards, including UALink and ESUN for scale-up networking and Ultra Ethernet for scale-out connections. For infrastructure buyers, the company's pitch is flexibility to combine networking with different accelerators as their computing requirements change.
An existing installation illustrates the work behind that pitch. In June, Cornelis said the Texas Advanced Computing Center accepted a CN5000 upgrade connecting more than 600 computing nodes in its Stampede3 supercomputer. Researchers use that system for applications including weather forecasting and engineering simulations. This provides a concrete deployment reference for Cornelis's networking business as it develops the new architecture.
SiliconANGLE identified IAG Capital Partners as the financing's lead investor. Cornelis says the money supports manufacturing, customer deployments and development of its next products. The company says CN5000 is already shipping, while CN6000 is being sampled by customers, with broader availability expected in the fourth quarter.
The next commercial test is translating the new architecture into working AI installations. Cornelis and Qualcomm are evaluating technology for rack-scale inference, according to Draeger's comments to SiliconANGLE. He also said a projected reduction in network traffic of up to 50% came from pre-production simulations. Whether those gains hold across customers' models and hardware will matter as the evaluation work progresses toward deployment.