Converge Digest

Video: How Cornelis Networks Is Rewiring AI Infrastructure

Check out other Tech Updates on our YouTube Channel (subscribe today): https://www.youtube.com/@NextGenInfra and check out our latest reports at: https://nextgeninfra.io/

In this interview, Cornelis Networks CEO Lisa Spelman outlines the company’s approach to building high-performance network infrastructure optimized for AI and HPC workloads. The discussion traces the origins of the Omni-Path architecture from its development at Intel to its evolution under Cornelis as an independent company.

Spelman explains how Cornelis has advanced the platform to address the performance limitations of traditional fabrics, focusing on predictable latency, lossless operation, and adaptive routing. She details the launch of the CN5000 series, a 400G end-to-end solution including custom superNICs, high-radix switches, and director-class systems designed for scale-out computing environments.

The conversation also highlights Cornelis’ silicon development, congestion management techniques, software ecosystem compatibility, and roadmap to 800G networking with convergence across Omni-Path and Ethernet.

This video provides a technical overview for those building or evaluating next-generation AI training infrastructure, exascale HPC systems, and large-scale GPU clusters.

⏱️ Key Topics & Timestamps:

00:00 – Introduction
01:13 – Limitations of legacy interconnects for large-scale parallel workloads
02:00 – Design goals behind the original Omni-Path architecture
04:00 – Transition from Intel to independent Cornelis Networks
05:00 – Adapting the architecture for AI training, inference, and HPC
05:40 – CN5000 platform components: superNIC, switch, and director-class systems
06:50 – Silicon design and first-pass tape-out success
08:00 – Credit-based flow control, adaptive routing, and congestion mitigation
09:30 – Scalability: supporting up to 500,000 connected GPUs
10:40 – Performance comparisons vs Infiniband and Ethernet at 400G
12:00 – Application-level improvements and GPU utilization gains
13:00 – Tuning the network for AI workloads and collective operations
14:00 – Commercial momentum and upgrade paths for existing customers
15:00 – Roadmap to 800G and Ethernet + Omni-Path convergence
16:00 – Multi-vendor support and compatibility with CUDA and ROCm
17:30 – Closing remarks

#AIInfrastructure #HPCNetworking #CornelisNetworks #OmniPath #400G #800G #DataCenterFabric #SiliconInterconnect #NetworkArchitecture #NextGenInfraHave a tech update that you want to brief us on? Contact info@nextgeninfra.io!
https://ngi.fyi/tech-cornelis-lisa

Exit mobile version