As AI infrastructure rapidly evolves, network performance has become just as important as compute power. Modern workloads such as large language model (LLM) training, generative AI, and high-performance computing (HPC) require ultra-low latency, massive bandwidth, and predictable communication between thousands of GPUs. As a result, many organizations are replacing legacy InfiniBand fabrics with next-generation 400G networking. The NVIDIA Quantum-2 MQM9790-NS2F is purpose-built for these environments. As a 64-port NDR 400Gb/s InfiniBand switch, it delivers exceptional throughput, advanced in-network computing, and the scalability required for AI factories, NVIDIA DGX SuperPOD deployments, and large HPC clusters. This guide explains the MQM9790-NS2F specifications, highlights its key advantages over previous-generation HDR switches, and outlines the most important technical and procurement considerations before deploying a Quantum-2 fabric.
Why Are AI Clusters Moving to Quantum-2?
As AI models continue to grow, communication between GPUs has become a major performance bottleneck. During distributed training, GPUs constantly exchange gradients and synchronization data. If the network cannot keep pace, expensive GPU resources remain idle while waiting for communication to complete.
Compared with previous-generation HDR InfiniBand, Quantum-2 significantly increases bandwidth while reducing latency and improving collective communication efficiency. These improvements enable organizations to:
- Accelerate distributed AI training
- Improve GPU utilization
- Reduce synchronization delays
- Scale clusters from hundreds to thousands of GPUs
- Support future AI infrastructure without redesigning the network
For enterprises investing in NVIDIA H100, H200, or Blackwell-based GPU platforms, Quantum-2 provides the networking foundation required for long-term scalability.
NVIDIA Quantum-2 MQM9790-NS2F Specifications
The MQM9790-NS2F is designed as a high-density leaf or spine switch for NDR InfiniBand fabrics.
Quick Specifications
| Specification | MQM9790-NS2F |
|---|---|
| Architecture | NVIDIA Quantum-2 |
| InfiniBand Speed | NDR 400Gb/s |
| Effective Port Count | 64 × NDR Ports |
| Physical Connectors | 32 × OSFP Twin-Port Cages |
| Switching Capacity | 51.2 Tbps |
| Form Factor | 1U |
| Management | Unmanaged |
| Airflow | Power-to-Connector (P2C) |
| Deployment Role | Leaf or Spine |
Check stock, compare options, or talk with our team.
High-Density 64-Port Design
Although commonly referred to as a 64-port NDR switch, the MQM9790-NS2F physically features 32 OSFP (Octal Small Form-factor Pluggable) cages. Each OSFP cage supports two independent 400Gb/s NDR connections through NVIDIA's twin-port design, providing an effective total of 64 NDR ports within a compact 1U chassis. This architecture delivers outstanding port density while minimizing rack space and simplifying large-scale fabric deployments.
51.2 Tbps Non-Blocking Switching Capacity
The Quantum-2 ASIC provides an aggregate switching capacity of 51.2 Tbps, allowing simultaneous communication between thousands of GPUs without creating network bottlenecks. The non-blocking architecture is especially beneficial for:
- Large Language Model training
- Deep learning workloads
- Distributed GPU communication
- Scientific simulation
- Massive MPI applications
Ultra-Low Latency Networking
Distributed AI applications rely on continuous synchronization across GPU nodes. Even small increases in latency can reduce overall cluster efficiency. Quantum-2 minimizes communication delays by delivering deterministic, ultra-low latency packet forwarding, enabling better scaling efficiency across large GPU environments.
Enhanced SHARP In-Network Computing
Quantum-2 enhances NVIDIA's Scalable Hierarchical Aggregation and Reduction Protocol (SHARP) technology. Instead of routing collective operations through CPUs or GPUs, SHARP performs reduction operations directly within the network, resulting in:
- Faster collective communication
- Reduced CPU overhead
- Higher GPU utilization
- Improved MPI performance
- Shorter AI training times
MQM9790-NS2F vs. Quantum HDR: What's New?
Organizations upgrading existing HPC environments often compare the MQM9790-NS2F with previous-generation Quantum HDR switches such as the MQM8790-HS2F.
| Specification | MQM9790-NS2F | MQM8790-HS2F |
|---|---|---|
| Architecture | Quantum-2 | Quantum HDR |
| Maximum Speed | 400Gb/s NDR | 200Gb/s HDR |
| Port Count | 64 | 40 |
| Switching Capacity | 51.2 Tbps | 16 Tbps |
| Connector Type | OSFP | QSFP56 |
| In-Network Computing | Enhanced SHARP | SHARP Supported |
| Typical Deployment | AI Factories, DGX SuperPOD | Traditional HPC Clusters |
The transition from HDR to Quantum-2 is more than a bandwidth upgrade. It also introduces a new physical connectivity standard, greater port density, and significantly improved scalability for AI-centric workloads.
Deployment Considerations
Understanding the Unmanaged Architecture
One important characteristic of the MQM9790-NS2F is that it is an unmanaged InfiniBand switch. Unlike managed Quantum-2 platforms, it does not include an onboard CPU capable of running the Subnet Manager (OpenSM). Instead, subnet management must be provided by:
- A dedicated host server
- An external OpenSM instance
- A managed Quantum-2 switch deployed within the same fabric
This design reduces hardware complexity while allowing the switch to function as a highly efficient forwarding device.
Typical Network Topology
The MQM9790-NS2F is commonly deployed in AI training clusters, NVIDIA DGX SuperPOD environments, enterprise AI infrastructure, and HPC research centers. Depending on the network architecture, it can serve as either a high-density leaf switch or part of a scalable spine-leaf fabric.
Its 64-port radix also enables architects to build large non-blocking Fat-Tree or DragonFly+ topologies with fewer switching tiers, reducing hop counts and overall latency.
Compatible Ecosystem
Selecting the switch is only one part of building an AI fabric. Compatibility across the entire networking stack is equally important.
| Component | Typical Recommendation |
|---|---|
| GPU Servers | NVIDIA DGX H100, DGX H200, DGX B200, HGX Platforms |
| Network Adapter | NVIDIA ConnectX-7 NDR |
| Optical Module | NDR OSFP Transceivers |
| Direct Attach Cable | NDR DAC Twin-Port Cables |
| Active Optical Cable | NDR AOC Twin-Port Cables |
| Fabric Software | OpenSM / NVIDIA UFM |
Building a reliable NDR InfiniBand fabric involves much more than selecting the right switch. Compatibility between Quantum-2 switches, ConnectX-7 adapters, OSFP transceivers, twin-port DAC/AOC cables, and GPU servers should be validated before deployment to minimize integration risks.
For enterprise AI projects, Router-switch.com's CCIE-certified engineering team can review your complete bill of materials (BOM), verify end-to-end compatibility, and help optimize the overall network design before procurement, reducing the likelihood of costly configuration changes during deployment.
Procurement Considerations
Purchasing an InfiniBand switch involves more than comparing hardware specifications. A successful deployment requires careful planning across network design, compatibility, and project scheduling.
Plan for Future Growth
AI clusters rarely remain static. Selecting a switching platform that supports future expansion reduces the need for disruptive infrastructure upgrades as GPU capacity increases.
Match Airflow to the Data Center
The MQM9790-NS2F uses Power-to-Connector (P2C) airflow. Before installation, verify that the airflow direction aligns with your hot-aisle and cold-aisle containment strategy, particularly in high-density AI racks.
Validate the Cabling Strategy
Because the switch uses 32 twin-port OSFP cages, organizations should carefully plan cable selection to match server connectivity requirements. Depending on deployment distance, suitable NDR DAC, AOC, or optical solutions should be selected to ensure optimal performance.
Consider Product Availability
AI infrastructure projects often operate under aggressive deployment schedules, where delays in networking equipment can postpone the entire cluster rollout—even if GPU servers are already installed.
Working with a trusted supplier that offers 100% genuine enterprise networking hardware, global warehouse inventory, and fast worldwide delivery can help keep deployment timelines on track. For organizations building multi-vendor AI infrastructures, Router-switch.com also provides competitive project pricing and coordinated sourcing for switches, optics, network adapters, and cabling to simplify procurement.
Conclusion
The NVIDIA Quantum-2 MQM9790-NS2F represents a major advancement in InfiniBand networking for AI and HPC. With 64 effective NDR 400Gb/s ports, 51.2 Tbps of non-blocking switching capacity, ultra-low latency, and enhanced SHARP in-network computing, it provides the performance and scalability required for modern AI factories, DGX SuperPOD deployments, and large-scale research environments.
Beyond switch performance, successful AI networking depends on selecting compatible components across the entire infrastructure—from GPU servers and ConnectX adapters to optical modules, cables, and fabric management software. Careful planning during the design and procurement stages helps reduce deployment risks while ensuring the network can scale with future AI workloads.
Whether you're building a new AI cluster or upgrading from Quantum HDR to Quantum-2, validating the complete solution before purchase is a best practice that can save significant time and cost throughout the project lifecycle.