Arrcus and QCT Forge Strategic Alliance for AI-Optimized Rack Solutions
Introduction: The Accelerating Need for Optimized AI Networking
The relentless advancement of Artificial Intelligence (AI) has placed unprecedented demands on underlying infrastructure, with networking emerging as a critical bottleneck and, conversely, a potential accelerant. Recognizing this pivotal role, Arrcus, a specialist in hyperscale networking software, and Quanta Cloud Technology (QCT), a prominent manufacturer of hyperscale hardware, have announced a strategic collaboration. This partnership aims to deliver integrated, AI-optimized rack solutions that combine Arrcus's robust ArcOS network operating system with QCT's high-performance switching platforms. The unveiled solution is designed to provide the high-bandwidth, low-latency connectivity essential for the most demanding AI workloads, from training massive models to deploying inferencing at the edge.
Arrcus and QCT: A Synergistic Partnership
This collaboration represents a significant move by both companies to address the burgeoning AI market. Arrcus, known for its hyperscale-grade networking software, brings its expertise in routing and switching infrastructure across core, edge, and multi-cloud environments. QCT, a global leader in hyperscale server and switch manufacturing, contributes its hardware prowess. By integrating Arrcus's ArcOS with QCT's latest switching platforms, the partnership promises a validated, rack-level solution specifically engineered for AI. Shekar Ayyar, Chairman and CEO of Arrcus, emphasized the synergy, stating, "As the AI revolution accelerates, networking has become the new accelerant. Together with Quanta, we are delivering a validated AI rack solution that combines our hyperscale-grade networking software with high-performance rack-scale QCT hardware to empower customers with agility, scale, and openness."
The ACE-AI Networking Stack: Powering AI Workloads
At the core of this integrated solution is Arrcus's ACE-AI, an intelligent networking stack meticulously engineered to meet the unique and stringent demands of AI workloads. This stack is built to support both flexible, scalable GPU cluster designs using IP CLOS architectures and Virtualized Distributed Routing (VDR). A key feature of ACE-AI is its ability to enable lossless Ethernet fabrics. This is achieved through the implementation of advanced protocols such as RoCEv2 (RDMA over Converged Ethernet v2), PFC (Priority Flow Control), and ECN (Explicit Congestion Notification). These technologies, combined with dynamic routing and sophisticated congestion control mechanisms, are crucial for ensuring high throughput and maintaining low latency across both AI training and inference processes. Furthermore, ACE-AI provides crucial hardware-level telemetry, offering real-time visibility into system health, buffer status, and performance counters. This detailed monitoring capability is designed for seamless integration with existing orchestration frameworks, providing operators with comprehensive control and insight.
Integrated Hardware Platforms and Key Technologies
As part of this strategic alliance, Arrcus is porting its ArcOS network operating system onto two of QCT's advanced, next-generation switching platforms. The first is the QuantaMesh TA064-IXM, an 800G switch designed for high-performance Leaf and Spine configurations within data center networks. The second is the QuantaMesh T1048-LYB, a 1G switch specifically designated for Out-of-Band (OOB) management, ensuring reliable control and monitoring access even under heavy network loads. This integration of software and hardware is crucial for delivering a cohesive and optimized solution.
Leveraging Broadcom Silicon for Enhanced Performance
The joint solution benefits significantly from the involvement of Broadcom, a leading provider of semiconductor solutions. Arrcus and QCT are collaborating to create a disaggregated, high-performance, and low-latency Ethernet backend network. This network leverages Broadcom's cutting-edge Tomahawk 5 silicon, which is a critical component for the demanding requirements of AI workloads. This architecture is designed to facilitate efficient and optimized data transfers across high-performance computing (HPC) clusters, AI applications, and storage systems. For the frontend network, the collaboration utilizes established merchant silicon, including both Broadcom's Tomahawk 5 and Trident silicon. This strategic use of advanced silicon enables the construction of a scalable Leaf-Spine fabric, providing a future-proof and robust foundation for environments driven by AI and HPC.
Targeted Use Cases and Benefits
The collaborative AI rack solution is meticulously designed to address several critical use cases within the AI ecosystem:
- AI Inferencing at the Edge: Delivering low-latency performance essential for real-time AI applications deployed at the network edge.
- Scalable AI Training Clusters: Providing the necessary bandwidth and low latency for training large, complex AI models in private and hyperscale data centers.
- Multitenant AI Fabrics: Enabling deterministic segmentation and isolation for environments hosting multiple AI tenants or workloads, ensuring security and performance predictability.
- Advanced Monitoring: Offering sophisticated application and network monitoring capabilities to ensure operational efficiency and rapid troubleshooting.
Mike Yang, Executive Vice President at Quanta Computer Inc. and President of QCT, highlighted the value proposition: "Partnering with Arrcus allows QCT to deliver high-bandwidth, low latency, scalable and validated AI rack solutions to our mutual customers. Together, we’re equipping next-generation data centers with cutting-edge switching platforms that support open networking standards to meet the demands of the AI era."
Industry Validation and Future Outlook
The collaboration has also garnered support from key industry players. Ram Velaga, senior vice president and general manager, Core Switching Group at Broadcom, commented on the integration of their technology: "Broadcom is proud to see our Tomahawk 5 Ethernet switching technology integrated into this joint solution from Arrcus and Quanta. High-performance, low-latency, and scalable networking is critical for AI workloads, and this collaboration showcases the value of combining open, innovative software with platforms built on Broadcom merchant silicon. Together, we’re enabling the next generation of AI infrastructure with the performance and flexibility customers demand."
Arrcus and QCT are not only showcasing this integrated solution but are also actively exploring further platform integrations and finalizing a comprehensive go-to-market strategy. This proactive approach aims to support customer deployments and ensure solution validation at scale. The partnership underscores a shared commitment to advancing open networking principles and collaboratively building the foundational infrastructure required for the next generation of AI advancements.
Conclusion: Paving the Way for AI Infrastructure
The strategic collaboration between Arrcus and QCT represents a significant development in the quest for optimized AI infrastructure. By combining specialized networking software with high-performance hardware, the partnership delivers a powerful, validated solution tailored to the unique demands of AI workloads. The focus on open networking, combined with cutting-edge silicon from Broadcom, positions this offering as a key enabler for data centers looking to harness the full potential of artificial intelligence. As AI continues its rapid evolution, such integrated and optimized solutions will be paramount in driving innovation and performance across the industry.
IP CLOS
IP CLOS (Cloud-native, Open, Scalable, Secure) is a network architecture commonly used in data centers. It is designed to provide high bandwidth and low latency, essential for modern applications like AI and High-Performance Computing (HPC). This architecture typically involves a multi-stage switching fabric, often with a spine-and-leaf design, ensuring that any server can connect to any other server with a predictable number of hops and high throughput. Its "open" nature often implies the use of standard protocols and merchant silicon, promoting flexibility and reducing vendor lock-in.
Virtualized Distributed Routing (VDR)
Virtualized Distributed Routing (VDR) is a networking approach that extends routing intelligence and functionality across multiple network devices, effectively creating a single, logical routing domain. Unlike traditional centralized routing where a few core routers manage the network
AI Summary
The collaboration between Arrcus and Quanta Cloud Technology (QCT) marks a significant step in addressing the escalating demands of Artificial Intelligence (AI) workloads by delivering optimized rack solutions. This strategic alliance integrates Arrcus's advanced ArcOS network operating system with QCT's cutting-edge switching platforms, creating a powerful combination of software and hardware designed to provide high-bandwidth, low-latency networking. The core of this integrated solution is Arrcus's ACE-AI intelligent networking stack, engineered to meet the specific challenges posed by AI applications. ACE-AI supports flexible and scalable GPU cluster designs through both IP CLOS and Virtualized Distributed Routing (VDR) architectures. It ensures high throughput and minimal latency for AI training and inference by enabling lossless Ethernet fabrics with technologies like RoCEv2, PFC, and ECN, coupled with dynamic routing and sophisticated congestion control mechanisms. Furthermore, the stack offers hardware-level telemetry for real-time system monitoring, seamlessly integrating with existing orchestration frameworks. As part of this collaboration, Arrcus will port its ArcOS onto two of QCT's next-generation switching platforms: the QuantaMesh TA064-IXM, an 800G switch suitable for both Leaf and Spine configurations, and the QuantaMesh T1048-LYB, a 1G switch intended for Out-of-Band (OOB) management. This integrated solution is specifically tailored for key AI use cases, including low-latency AI inferencing at the edge, the deployment of large-scale AI training clusters in private and hyperscale data centers, and the creation of multitenant AI fabrics that offer deterministic segmentation and isolation. It also enhances application and network monitoring capabilities. The partnership, which also involves Broadcom, focuses on creating a disaggregated, high-performance, low-latency Ethernet backend network utilizing Broadcom's Tomahawk 5 silicon, crucial for efficient data transfers in AI, High-Performance Computing (HPC), and storage environments. The frontend network will leverage existing merchant silicon, including Broadcom's Tomahawk 5 and Trident series, to build a scalable Leaf-Spine fabric, ensuring a future-proof infrastructure for AI and HPC demands. Both companies are committed to advancing open networking principles and are exploring further platform integrations and go-to-market strategies to support broad customer adoption and large-scale solution validation. This initiative underscores their shared vision of building the next generation of AI-ready infrastructure.