Cornelis Networks and the Active Compute Fabric
Cornelis Networks, a company with deep roots in high-performance computing, is making a aggressive push into the AI scale-up market with the introduction of its Active Compute Fabric (ACF). By establishing an open architecture, Cornelis aims to provide a standardized approach to both scale-up and scale-out networking. This move is designed to integrate programmable compute directly into the fabric, effectively decentralizing the tasks typically handled solely by GPUs.
The technical ambition here is to mirror the efficiency of Nvidia’s proprietary SHARP technology, which offloads collective operations to network switches. Cornelis’ ACF goes beyond simple data movement by offloading specific, high-overhead AI tasks. Key features include KV cache offloading to streamline state management, Message Passing Interface (MPI) offloading for HPC workloads, and in-fabric checkpointing to improve failure recovery. By moving these processes into the network, Cornelis claims it can reclaim vast amounts of idle GPU time, noting that in massive 100,000-GPU clusters, nearly half of all compute capacity is currently squandered while waiting for data synchronization.
Delos Data: Bridging the Physical Gap with Protocol Agnosticism
While Cornelis focuses on architectural standards, Delos Data is tackling the physical limitations of current AI hardware. Founded by veterans from Barefoot Networks and Intel, Delos has unveiled its Nonstop AI portfolio, a series of hardware designs aimed at scaling AI beyond the constraints of a single rack. The core of their strategy is protocol agnosticism; their I/O chiplets are designed to function regardless of whether the customer uses Ultra Accelerator Link (UALink), Ethernet for Scale-Up Networking (ESUN), or other emerging standards.
Delos is offering three distinct, high-performance form factors. Their I/O die is the most ambitious, boasting over 30 Tbps of aggregate bandwidth—a massive leap over the 3.6 TB/s limits currently seen in leading accelerator interconnects. For connectivity between racks, the company has developed Near-Packaged Optics (NPO) that deliver 10 Tbps of interface bandwidth, allowing for greater scalability without the power-draw limitations of traditional copper. Complementing this is a new 400+ Gbps NIC designed to serve as the backbone for large-scale training clusters and front-end storage access, all managed by the Nonstop AI software platform for dynamic traffic rerouting.
Why It Matters
- Breaking the Walled Garden: By championing open standards, these firms provide hyperscalers with alternatives to Nvidia's proprietary NVLink ecosystem, potentially lowering costs and hardware lock-in.
- Energy Efficiency: Reducing the time GPUs spend in idle states waiting for data movement could save hundreds of gigawatt-hours of power in massive data centers annually.
- Scalability: Current systems are limited by physical copper constraints; the adoption of near-packaged optics and specialized I/O chiplets is essential for scaling from 72-GPU racks to clusters of over 1,000 GPUs.
The influx of venture capital into both firms—with Cornelis securing $205 million and Delos raising over $100 million—underscores the industry’s hunger for a viable, competitive alternative to Nvidia’s vertically integrated stack. As these technologies move from design to deployment, the focus will shift to how effectively these hardware-agnostic solutions can perform in real-world, large-scale AI training environments.
