Cornelis and NextSilicon announced a collaboration to develop and evaluate joint reference architectures for artificial‑intelligence (AI) and high‑performance computing (HPC) systems. The partnership pairs Cornelis’s CN5000 400 Gbps fabric with NextSilicon’s Maverick‑2 dataflow accelerator, aiming to reduce network‑induced idle time and improve compute utilization for enterprise‑grade workloads. By targeting the two most common sources of inefficiency—network congestion and von Neumann‑style compute stalls—the alliance promises a more balanced, end‑to‑end solution that can be adopted by original‑equipment manufacturers (OEMs) without requiring customers to redesign their software stacks.
Joint Reference Architecture Initiative
At ISC High Performance 2026, the two companies detailed a program that will produce validated system blueprints for OEMs. The first phase combines the CN5000 fabric—launched in 2025—with Maverick‑2, which entered volume production late 2025. Joint testing focuses on how the fabric and accelerator perform together across multiple configurations, giving OEM partners a proven starting point rather than an untested parts list. The roadmap includes extending the evaluation to the 800 Gbps CN6000 fabric, slated for release in the second half of 2026, and later exploring emerging AI workloads such as disaggregated inference and agentic AI. By delivering these blueprints early, Cornelis and NextSilicon aim to shorten time‑to‑market for next‑generation AI/HPC systems and provide a clear migration path from legacy Ethernet‑based designs.
Technical Approach to Network and Compute Bottlenecks
Cornelis targets the latency‑sensitive, small‑message traffic typical of AI inference and HPC simulation, which standard Ethernet struggles to handle without congestion. The CN5000 is described as “congestion‑free,” intended to keep data moving and prevent expensive compute resources from waiting on the network. On the compute side, NextSilicon moves away from the traditional von Neumann architecture. Its Maverick‑2 accelerator implements an Intelligent Compute Architecture (ICA) that reconfigures at runtime to match workload characteristics, while still executing existing code without modification. By pairing a fabric that eliminates network stalls with a dataflow accelerator that adapts to software demands, the joint reference designs aim to deliver a more efficient AI/HPC foundation that maximizes hardware utilization and reduces overall system cost.
Enterprise Relevance and Early Feedback
Lisa Spelman, CEO of Cornelis, noted that “operators keep telling us their most expensive systems sit idle, waiting on the network,” underscoring the commercial pain point the collaboration seeks to alleviate. Elad Raz, founder and CEO of NextSilicon, added that Maverick‑2 “makes the processor adapt to the software,” positioning the accelerator as a counterpoint to the long‑standing practice of forcing software to fit the processor. Both executives framed the partnership as a natural fit to demonstrate a “congestion‑free fabric and a workload‑driven compute architecture” to customers and OEM partners.
Key Takeaways
- The collaboration pairs Cornelis’s 400 Gbps CN5000 fabric (launched 2025) with NextSilicon’s Maverick‑2 accelerator (volume shipping late 2025).
- Joint evaluations will produce validated AI/HPC reference architectures for OEM partners, with plans to test the upcoming 800 Gbps CN6000 fabric in late‑2026.
- Both companies cite network congestion and von Neumann compute stalls as primary inefficiencies that their combined solution intends to mitigate.
TechInsyte's Take
The joint effort offers a concrete blueprint for enterprises seeking to maximize utilization of high‑cost AI and HPC hardware. While the announced designs address known latency and compute bottlenecks, broader adoption will depend on OEM integration timelines and real‑world performance data from the upcoming CN6000 fabric tests. Decision‑makers should monitor the validation results and OEM roadmaps to assess fit for their own AI/HPC deployments.
Source: Cornelis
