UfiSpace and Arrcus Deliver 102.4T AI Networking Architecture with Direct Liquid-Cooling for AMD-Based AI Data Centers
Integrated open networking solution leverages UfiSpace 102.4T/51.2T hardware platforms and Arrcus ArcOS® to optimize GPU-driven RoCEv2 traffic across AMD Instinct™ GPU clusters, reduce Job Completion Times, and resolve AI cluster power constraints.
SAN FRANCISCO, CA — July 23rd, 2026 — UfiSpace, a global leader in open networking and disaggregated communications solutions, today announced an expanded strategic collaboration with Arrcus, the hyperscale networking software company. Demonstrated at AMD Advancing AI 2026, the joint solution delivers a comprehensive, production-ready disaggregated networking stack engineered to support AMD-powered, high-density, multi-terabit AI data center fabrics.
As enterprise AI clusters scale to tens of thousands of accelerators, the network fabric becomes the primary determinant of system-wide efficiency. AI training workloads generate massive East-West traffic driven by GPUs during collective communication. To prevent network congestion and packet loss, the UfiSpace and Arrcus architecture pairs high-density 1.6T/800G white-box platforms with the Arrcus ArcOS® network operating system, providing loss-sensitive RoCEv2 transport, deterministic latency, and optimized Total Cost of Ownership (TCO).
Addressing Thermal Density and Scale-Out AI Bottlenecks
Modern hyperscale AI infrastructure, increasingly built on AMD Instinct™ GPU platforms, demands the seamless integration of high switching capacity, advanced thermal management, and compliance with open specifications.
"Scaling next-generation AI clusters requires a system-level hardware design capable of aligning switching capacity, thermal engineering, and software agility," said Vincent Ho, CEO of UfiSpace. "Our strategic alignment with Arrcus demonstrates the production readiness of disaggregated networking. By leveraging UfiSpace's deep engineering expertise in 102.4 Tbps platforms and direct liquid-cooling architectures alongside Arrcus ArcOS®, we empower operators to deploy open, scalable AI fabrics that optimize Job Completion Times while addressing stringent power and density requirements."
"Open networking is essential for scaling AI infrastructure efficiently and sustainably," said Shekar Ayyar, Chairman and CEO of Arrcus. "Integrating Arrcus ArcOS® with UfiSpace’s multi-terabit hardware platforms gives operators building on AMD-powered AI infrastructure a solution that resolves critical bandwidth and congestion bottlenecks. Together, we provide advanced routing, robust congestion control, and high-availability fabrics engineered for large-scale AI workloads."
Key Technical & Solution Highlights
● 102.4 Tbps Ultra-High-Density Fabric with Liquid-Cooling Support:
Featuring the S9331-64HO and S9331-64HO3L platforms powered by Broadcom Tomahawk® 6 ASICs, UfiSpace delivers an industry-leading 102.4 Tbps of switching capacity. Designed specifically for next-generation 1.6T/800G scale-out AI clusters, these platforms offer Direct Liquid-Cooling options to manage extreme thermal densities, allowing operators to optimize Power Usage Effectiveness and maintain peak performance under continuous AI workloads.
● 51.2 Tbps High-Density Scale-Out Fabric:
The S9321-64EO platform offers 51.2 Tbps switching capacity powered by Broadcom Tomahawk® 5 silicon in a compact 2RU footprint. Delivering 64 ports of 800G/400G QSFP-DD density, it provides predictable low latency and optimized RoCEv2 traffic handling to minimize packet loss and slash AI Job Completion Times (JCT).
● GPU-Driven Workload Optimization & End-to-End Portfolio Readiness:
Engineered to absorb intensive GPU-driven East-West traffic patterns, the joint solution was validated in environment configurations leveraging AMD Instinct™ GPU platforms and AMD EPYC™ processors. The architecture spans Spine, Leaf, Access, and Out-of-Band (OOB) management tiers, incorporating UfiSpace S9300-32D (Trident 4) and S6301-56ST (Trident 3) platforms, all seamlessly controlled via Arrcus ArcOS® to ensure end-to-end fabric interoperability.
● Data Center-Scale Congestion Control & Software Disaggregation:
Arrcus ArcOS® provides granular congestion management, including Priority-based Flow Control (PFC) and Explicit Congestion Notification (ECN), to eliminate packet drops in loss-sensitive AI fabrics. The microservices-based architecture enables advanced L2/L3 routing, real-time telemetry, and automated fabric deployment without proprietary hardware constraints.
About UfiSpace
UfiSpace is a leading provider of open networking solutions, dedicated to driving cost efficiency, flexibility, and technological innovation across telecom carriers, cloud service providers, and hyperscale data centers. Through a comprehensive portfolio of disaggregated white-box switches, routers, and transport platforms, UfiSpace empowers global network operators to build open, scalable, and future-proof infrastructures. For more information, visit the official website.
Follow us on our social media channels: LinkedIn, Facebook, and YouTube
About Arrcus
Arrcus is a leading provider of high-performance routing and switching solutions, enabling organizations to achieve superior scalability and reliability. Its ACE platform, powered by the ArcOS network operating system, offers an agile alternative to traditional monolithic systems for datacenter, core, and edge environments.
To meet the distinct demands of AI, Arrcus delivers specialized ethernet fabrics. For training, its scale-out and scale-across solutions handle massive data synchronization and high-bandwidth requirements. For inferencing, the company recently launched the Arrcus Inference Network Fabric (AINF), a policy-driven approach spanning edge, datacenter, and cloud. By reducing token latency and improving throughput, AINF lowers the cost per token and enables sovereign AI, helping global enterprises modernize connectivity for the AI-driven era. The company is headquartered in San Jose, California. For more information, visit the official website or follow Arrcus on LinkedIn and Twitter/X.
MEDIA CONTACT
Name: Betty Wu
E-mail: betty.wu@ufispace.com