NVIDIA Quantum-2 QM9700 QM9790 Ethernet Switch

The NVIDIA Quantum-2 QM9700 QM9790 Ethernet Switch stands at the forefront of high-performance computing (HPC) and artificial intelligence (AI) networking. Built to power the most demanding research, innovation, and product development, this switch system delivers an unparalleled combination of ultra-low latency, massive bandwidth, and scalable fabric capabilities. It is engineered to unify compute, storage, and accelerator resources into a single, deterministic network that accelerates communications between GPU-enabled nodes, accelerates data movement, and simplifies data-center management. With Quantum-2, data centers can push the boundaries of what is possible in simulations, model Training, and real-time analytics, all while maintaining reliability, security, and ease of operation across large-scale deployments.

• Unparalleled performance for AI and HPC workloads: Sub-microsecond latency and line-rate throughput empower GPU-to-GPU and node-to-node communications, ensuring accelerators remain fed with data and workloads stay in the compute zone. The Quantum-2 fabric is designed to minimize jitter and deliver deterministic results, enabling tightly coupled simulations, multi-node neural network training, and latency-sensitive inference at scale. This performance translates into faster time-to-insight and more efficient use of expensive GPU resources across clusters, racks, and data centers.
• Scalable, high-density fabric: Engineered for growth, the QM9700 QM9790 family supports flexible port configurations and breakout options to accommodate evolving workloads. As data-center needs expand—from modest clusters to multi-rack deployments—the switch scales without compromising performance or manageability. Built with modular, hot-swappable components and efficient cooling, it maintains stable operation in dense environments while simplifying capacity planning and expansion.
• RDMA and RoCE v2 support: Remote Direct Memory Access (RDMA) and RoCE v2 reduce CPU overhead by enabling direct, low-latency memory-to-memory transfers across the network. This capability is essential for GPU-accelerated workloads, large-scale distributed training, and data-intensive analytics where CPU involvement can become a bottleneck. RDMA helps achieve higher effective bandwidth, reduced latency, and more predictable performance under load.
• Programmable and standards-based: The Quantum-2 switch embraces open standards and programmable interfaces, enabling seamless integration with orchestration tools, monitoring platforms, and NVIDIA software stacks. With API-driven automation, you can implement repeatable configurations, rapid deployment of fabric policies, and consistent performance across clusters. This approach reduces vendor lock-in and accelerates the adoption of modern, software-defined networking practices in research and production environments.
• Seamless NVIDIA AI ecosystem integration: Optimized for CUDA and NVIDIA software libraries, the Quantum-2 switch accelerates distributed training, model serving, and data movement across the AI pipeline. It complements NVIDIA’s AI software stack, enabling efficient multi-node training, data-parallel and model-parallel strategies, and faster experiment cycles. The result is a more productive AI workflow—from data ingestion and preprocessing to training and inference—within a unified, scalable fabric.

Technical Details of NVIDIA Quantum-2 QM9700 QM9790 Ethernet Switch

• Model family: QM9700 and QM9790 Ethernet Switch variants designed for HPC and AI networks
• Form factor: 2U rack-mountable data center appliance suitable for dense compute environments
• Port configurations: high-density options with flexible breakout and multi-rate support (port counts and speeds vary by SKU)
• Switching fabric: non-blocking, deterministic fabric engineered for predictable, low-latency interconnects across thousands of nodes
• Latency: sub-microsecond class performance to meet the most demanding time-sensitive workloads
• Management: centralized, secure management with role-based access control and API-driven automation for scalable operations
• Power and redundancy: redundant, hot-swappable power supplies and fan modules for mission-critical reliability
• Security: hardware-assisted protections and encryption options to safeguard data in transit across the fabric
• Network protocols and features: RoCEv2, RDMA capabilities, NVMe over Fabrics readiness, QoS, traffic shaping, and robust fabric provisioning

Specifications may vary by SKU. For exact port counts, speeds, and features, refer to the Synnex listing for the corresponding UPC or SKU.

How to install NVIDIA Quantum-2 QM9700 QM9790 Ethernet Switch

• Plan your topology and prepare the data center environment: verify rack space (2U form factor), cooling capacity, and redundant power availability. Map out the fabric architecture, including spine/leaf topology if applicable, and determine breakout configurations to optimize workload placement.
• Mount the switch in a standard 19-inch rack, ensuring clear airflow and accessibility to management interfaces. Verify that cabling paths are organized to minimize interference and ease maintenance, with attention to cable length limits and bend radii for high-speed optics.
• Connect fabric links to compute nodes and storage endpoints using recommended multi-rate ports. Use appropriate cabling (e.g., 400G QSFP28 or equivalent) and configure breakout groups if you’re using multi-rail or multi-chassis topologies to balance bandwidth across workloads.
• Power up and perform initial configuration: access the management interface, apply network policies, configure QoS, VLANs, routing (if applicable), and fabric-level settings. Establish baseline security configurations, including user roles and access controls, to protect the fabric from unauthorized changes.
• Integrate with orchestration and monitoring tools: register the switch with your data-center fabric manager, enable automation workflows, and implement health checks. Validate performance with representative workloads, monitor latency and bandwidth, and tune QoS and traffic policies to align with your HPC or AI priorities.

Frequently asked questions

• What workloads is the NVIDIA Quantum-2 QM9700 QM9790 best suited for? The Quantum-2 Ethernet Switch is designed for large-scale HPC simulations, AI training and inference at scale, data analytics across distributed clusters, and any workload that benefits from low latency, high throughput interconnects and deterministic performance across thousands of nodes.
• Does it support RDMA and NVMe over Fabrics? Yes, the Quantum-2 family provides RDMA capabilities (RoCEv2) and is designed to work with NVMe over Fabrics in appropriate configurations to accelerate storage traffic and reduce CPU overhead during data transfers.
• How scalable is the Quantum-2 switch? It is built for large-scale deployments, offering flexible port configurations and the ability to expand across multi-rack fabrics to accommodate growing compute resources and data demands without compromising performance or manageability.
• What management options are available? The switch offers centralized, secure management with API-driven automation, role-based access control, firmware and policy updates, and integration with common orchestration and monitoring platforms to simplify fabric provisioning and ongoing operations.
• What are the key benefits for NVIDIA AI workflows? The Quantum-2 switch reduces communication bottlenecks in multi-GPU training, enables faster gradients exchange, improves data movement across the AI cluster, and supports a cohesive software ecosystem that accelerates the overall experimentation cycle.