Arrcus and UfiSpace introduced a production-ready AI networking solution that combines the Arrcus ArcOS network operating system with UfiSpace’s Broadcom-based open Ethernet switches for AMD-powered AI infrastructure. The integrated platform, demonstrated at AMD Advancing AI 2026 in San Francisco, is designed to support large-scale AI training and inference using AMD EPYC CPUs and AMD Instinct accelerators deployed within GIGABYTE’s modular GIGAPOD AI cluster architecture.
The jointly validated solution brings together Arrcus’ ArcOS software and UfiSpace’s open networking hardware based on Broadcom Tomahawk 6, Tomahawk 5, Trident 4, and Trident 3 switching silicon. The companies position the platform as an open alternative to proprietary AI networking fabrics, supporting lossless RoCEv2 Ethernet, congestion management, advanced load balancing, deep telemetry, and multi-tenant operation. UfiSpace contributes high-density switch platforms, including liquid-cooled, OCP ORv3-ready systems intended for hyperscale AI deployments, while Giga Computing integrates the networking into complete AMD-based AI clusters.
The reference architecture includes switching platforms ranging from 102.4 Tbps Tomahawk 6 systems for AI fabrics to lower-speed Trident-based switches for storage, aggregation, server access, and out-of-band networking. According to the companies, ArcOS is co-validated and performance-tuned on each hardware platform to reduce deployment complexity while optimizing traffic flows for AMD Instinct-based AI clusters.
| AI Networking Platform Highlights | |
| Network OS | Arrcus ArcOS with RoCEv2, congestion management, telemetry and load balancing |
| Switch Silicon | Broadcom Tomahawk 6, Tomahawk 5, Trident 4 and Trident 3 |
| Maximum Fabric Capacity | 102.4 Tbps (Tomahawk 6 platforms) |
| Spine Platform | S9321-64EO with Tomahawk 5 delivering up to 51.2 Tbps |
| AI Platform | AMD EPYC processors and AMD Instinct accelerators integrated into GIGAPOD clusters |
| Deployment Features | Open Ethernet fabric, co-validated hardware/software stack, liquid-cooled switch options, OCP ORv3-ready platforms |
| Target Workloads | Large-scale AI training and inference clusters |
“By integrating ArcOS deeply with UfiSpace’s Tomahawk 6 and Tomahawk 5 Broadcom platforms, we give customers building AMD-powered AI clusters an open, high-performance Ethernet fabric they can deploy with confidence,” said Shekar Ayyar, chairman and CEO of Arrcus.
🌐 Analysis
The announcement reflects the growing momentum behind Ethernet-based AI networking as hyperscalers and enterprise infrastructure providers seek alternatives to proprietary AI interconnects. Rather than introducing a new networking technology, Arrcus and UfiSpace emphasize software validation and system integration around Broadcom’s latest Ethernet switching silicon and AMD compute platforms, reducing the engineering effort required to deploy production AI clusters.
The partnership also illustrates the expanding ecosystem around AMD’s AI infrastructure strategy. Alongside AMD processors and accelerators, the reference architecture combines networking from Arrcus and UfiSpace with Giga Computing’s GIGAPOD systems and Broadcom switching silicon, providing customers with an integrated, multi-vendor Ethernet solution for AI training and inference.
