Arrcus and UfiSpace launch open Ethernet fabric for AMD AI clusters
Arrcus and UfiSpace have announced the integration of the ArcOS network operating system with Broadcom-based open switches to support high-performance AMD GPU and CPU clusters. This pre-validated networking fabric is designed for large-scale AI training and inference workloads and was showcased alongside GIGABYTE's modular AI solutions at AMD Advancing AI Day 2026.
Key Takeaways
- The 102.4 Tbps S9331-64HO switch utilizes Broadcom Tomahawk 6 silicon to support high-radix AI fabrics.
- ArcOS provides RoCEv2 lossless-Ethernet, congestion management, and deep telemetry optimized for AMD Instinct-based clusters.
- The integrated solution serves as the foundation for GIGABYTE's modular 'GIGAPOD' AI rack solutions.
- Hardware options include liquid-cooled, OCP ORv3-ready platforms designed for next-generation power density.
Why It Matters
This partnership provides an alternative to proprietary networking fabrics like NVIDIA's InfiniBand, offering an open Ethernet-based pipeline for AMD-centric data centers. By co-validating software and hardware, Arrcus and UfiSpace reduce the integration risk for operators scaling 100T-class networks for generative AI. This move strengthens the ecosystem surrounding AMD's hardware as it competes for the #2 spot in the AI accelerator market. Watch for volume adoption rates of the Tomahawk 6 platform versus NVIDIA's Spectrum-X1600 Ethernet switches, expected to hit high-volume production in late 2026.
Additional Context
The rollout of the 102.4 Tbps Tomahawk 6 platform is a central component of Broadcom's strategy to capture the 'networking phase' of the AI cycle. Per Broadcom and StockTitan (March 2026), the company achieved production-volume shipments of the Tomahawk 6 in under nine months from initial sampling, significantly outpacing its historical development cycles. This silicon supports up to 512 channels of 200G SerDes, providing the necessary bandwidth to interconnect clusters of up to one million AI accelerators, which is becoming a baseline requirement for training trillion-parameter models like Llama-4. While NVIDIA continues to dominate the AI accelerator market with an estimated 80% share, AMD's Instinct line is positioning itself as a credible open-ecosystem alternative. Per Silicon Analysts (April 2026), AMD’s MI350X matches or exceeds NVIDIA's B200 in raw memory capacity, though it faces ongoing challenges in software-driven performance utilization. These open networking collaborations are critical to bridging that gap; per Dell’Oro Group (June 2026), Ethernet switch sales in AI back-end networks more than doubled in early 2026, now accounting for two-thirds of all AI cluster switch sales as enterprises push back against proprietary vendor lock-in. Arrcus is also leveraging this infrastructure for its recently launched Inference Network Fabric (AINF). Per Futuriom (February 2026), Arrcus reported a 300% booking growth in 2025, driven largely by demand for distributed AI inference solutions. The AINF platform uses ArcOS's policy-aware routing to steer traffic between edge nodes and data centers, optimizing for latency and power efficiency. By integrating with UfiSpace's 1.6T-ready portfolio, Arrcus aims to support the transition toward 'Agentic AI' workloads that require faster token retrieval times than traditional data center fabrics can provide.
Read full article at google.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source