Cornelis raises $205 million to put computing into the AI cluster network

Intel's former networking branch introduces Active Compute Fabric, which embeds computing functions in network silicon to reduce GPU waiting.

Cornelis has completed a $205 million round led by IAG Capital Partners, and in the wake of it unveiled Active Compute Fabric. The principle: instead of simply connecting nodes, the network embeds programmable functions directly in its silicon, so that data can be processed and transmitted simultaneously.

The problem is trivial but costly: in a training cluster, a significant part of the GPU time is spent waiting for data to arrive. The company announces that it will be able to offload collective operations, adapt the behavior of the network to the load and add functions as AI software evolves. The launch is accompanied by a collaboration with Qualcomm.

Born out of Intel in 2020, Cornelis relies on open architecture to distinguish itself from Nvidia: its customers can combine the network fabric with GPUs and accelerators from various vendors, whereas InfiniBand and NVLink assume the manufacturer's ecosystem. Its 400 Gbit/s generation is in production, the next 800 Gbit/s is expected.