Data center infrastructure startup Cornelis Networks Inc. is taking a major leap forward with the launch of a new Active Compute Fabric architecture to support scale-up and scale-out networks for artificial intelligence workloads.
The new fabric was announced alongside a strategy collaboration with Qualcomm Technologies Inc. and a $205 million funding round led by IAG Capital Partners.
Cornelis is a developer of specialized and congestion-free data center networking systems for AI and high-performance computing. A rival to more established providers such as Cisco Systems Inc. and Arista Networks Inc., its network infrastructure is designed to accelerate those workloads and maximize compute performance to enhance AI model training and inference.
The Active Compute Fabric builds on that foundation. Announced today at the AI Infra Summit, it’s a new, open architecture that integrates programmable compute directly into the fabric of the network, making it far more flexible and efficient. Cornelis said the rapid growth in AI model compute cluster sizes has created some serious bottlenecks in data center networks, leaving expensive AI accelerators sitting idle for far too long while they wait for the data they need to process.
In the last couple of years, compute, memory and storage have all evolved to become more “workload-aware,” but traditional networks have not yet undergone the same change, which means they’re struggling to keep up, said Cornelis Chief Executive Lisa Spelman.
Active Compute Fabric is meant to change this, It combines in-fabric acceleration with lossless transport and programmable compute so that the network can actively work with data even as it traverses the system. Because of this, it can adapt in real time based on the shifting workloads it’s asked to handle, offload complex collective operations and take on new functions. Even better, it’s built on open standards such as Ethernet and UALink for scale-up and Ultra Ethernet for scale-out, so that organizations can implement it with their existing compute architectures.
Cornelis Chief Marketing Officer Brandon Draeger told SiliconANGLE that the Active Compute Fabric performs certain workload-specific operations as data moves through the network, rather than just acting as a kind of messenger. For instance, it can assemble KV cache data for disaggregated inference, coordinate expert dispatch for mixture-of-experts models and accelerate collective operations like AllReduce and compress gradients in transit. All this is done before the data arrives at its destination.
“The payload does not arrive the way it left,” Draeger said. “In a collective operation, partial results from many endpoints are combined inside the fabric, so a single reduced result lands at the destination instead of thousands of separate contributions. Compressed gradients move as a fraction of their original size. The work happens once, in the path the data was already taking, instead of consuming accelerator cycles at both ends.”
This provides two key advantages. It means that less data has to traverse the network, and it means the GPUs spend less time waiting on communications and on synchronizing the information that arrives. According to Draeger, this helps to reduce overall network traffic by up to 50%, based on findings from Cornelis’s pre-production simulations.
“Accelerator utilization in large AI deployments commonly sits near half of installed capacity, and the architecture is designed to return a meaningful share of what is currently wasted,” he explained. “How much depends on the model, the cluster size, and the customer’s stack, so improvements will vary by workload and deployment.”
Spelman said Cornelis is expanding into scale-up and scale-out so it can bring its network architecture closer to the AI accelerators that power AI workloads in order to feed them with data faster and enhance their performance. “AI infrastructure is reaching a point where faster endpoints alone are not enough,” she said. “The fabric has to become an active part of the compute system. We’re seeing growing demand from customers for an open alternative that gives them more choice in how they build their AI infrastructure, and that approach is creating real momentum for Cornelis.”
Cornelis’ vision of open and efficient scale-up and scale-out networks is shared by Qualcomm, which will join it on stage at the AI Infra Summit in Santa Clara to announce a strategic collaboration aimed at shaping the future of networking for rack-scale AI data centers. They both see existing, passive network fabrics that only move data as a massive problem that causes AI accelerators to be constantly underutilized, and believe that networks should be given more priority in the design of AI data center infrastructures.
Draeger explained that this happens because existing data center networks are like “passive pipes” that simply just forward data packets and do nothing else. “It means every collective, every synchronization step and every cache transfer costs accelerator time,” he said.
At present, the network represents about 15% of the cost of an AI system, but it also determines how much value you get out of the other 85% of what was spent. Cornelis and Qualcomm believe the network should be treated as a “first-order design decision,” which means building a scale-up and scale-out network alongside the accelerator, not after the compute decision is made.
“Qualcomm and Cornelis share this view, and we are in advanced stages of joint technology evaluation focused on rack-scale inference across both scale-up and scale-out environments,” Draeger said. “The goal is a system where the network and the accelerators are designed to work together from the start.”
Going forward, Cornelis will tap the $205 million in new capital it has raised to scale the production of its CN5000 and CN6000 network switches and accelerate the deployment of its Active Compute Fabric.
