en.Wedoany.com Reported - Cornelis Networks has partnered with chip startup NextSilicon to develop a reference architecture for AI and high-performance computing (HPC), aiming to address bottleneck issues in standard Ethernet under latency-sensitive burst traffic scenarios.

This collaboration will integrate Cornelis' CN5000 switching fabric with NextSilicon's Maverick-2 computing platform. Both parties believe that standard Ethernet is unsuitable for AI inference and HPC simulation workloads, as the burst traffic it generates can easily lead to congestion. To address this, the reference architecture adopts a switching fabric that maintains data flow and an accelerator that keeps computing busy.

Cornelis' CN5000 has been deployed in several large-scale systems, including the Lynx supercomputer at Lawrence Livermore National Laboratory (LLNL), which was officially deployed last week. Additionally, the technology is used in the "Genesis Mission," an initiative driven by the Trump administration aimed at digitizing vast amounts of historical scientific data from tape systems for use by AI models.

NextSilicon's Maverick-2 Intelligent Compute Accelerator (ICA) chip, released in October 2024, is manufactured using Taiwan Semiconductor Manufacturing Company's (TSMC) 5-nanometer process. The hardware claims to feature a "novel and original computing architecture" and has already secured several HPC customers, including serving the U.S. Department of Energy through Sandia National Laboratories' Spectra supercomputer.

Both parties assert that their integrated stack will overcome infrastructure bottlenecks that have previously led to underutilization of expensive systems. In the first phase of the collaboration, they will validate the synergistic performance of the switching fabric and computing in various configurations, enabling original equipment manufacturer (OEM) partners to start from a proven combination. The project will later expand to higher network speeds, utilizing the upcoming 800 gigabits per second (Gb/s) CN6000 switching fabric.

Lisa Spelman, CEO of Cornelis Networks, stated that operators report the most expensive systems sit idle while waiting for network connections. The company's CN5000 was developed precisely to solve this problem. NextSilicon also challenges traditional assumptions in computing, making the two companies' collaboration complementary.

Elad Raz, founder and CEO of NextSilicon, noted that over the past few decades, software had to adapt to processors, but Maverick-2 makes the processor adapt to software, and Cornelis takes a similar approach to networking.

Additionally, Cornelis Networks revealed that its CN5000 network technology has been deployed as part of the Stampede3 super server upgrade. This supercomputer, located at the Texas Advanced Computing Center (TACC), is based on Dell and Intel processors and is a 10 petaflop system used for simulation workloads. The CN5000 deployment covers over 600 compute nodes, improving performance and scalability, enabling simulation workloads such as weather forecasting, engineering, and data-intensive analysis to complete faster. Cornelis Networks stated it will continue to support TACC's mission, helping researchers tackle increasingly complex scientific challenges.

This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com