NVIDIA MQM9790-NS2F MELLANOX Quantum 2 Based NDR Infiniband

NVIDIA MQM9790-NS2F Quantum 2 NDR InfiniBand Switch

Overview

The NVIDIA MQM9790-NS2F is a Quantum 2-based NDR InfiniBand switch designed for next-generation data center fabric, high-performance computing clusters, and storage-dense infrastructure. This is not a conventional network switch—it is purpose-built for environments where ultra-low latency and massive parallel data transfer are non-negotiable. The MQM9790-NS2F (often searched as MQM9790 NS2F) integrates NVIDIA's proven InfiniBand architecture to deliver 400Gbps port throughput, making it the backbone choice for GPU compute clusters, AI training pipelines, and scientific computing deployments where traditional Ethernet introduces unacceptable bottlenecks.

Key Features

• NDR (Next Data Rate) InfiniBand Architecture: 400Gbps per-port throughput eliminates the bandwidth ceiling that constrains multi-GPU systems. When training large language models or running distributed simulations across dozens of accelerators, a single oversubscribed Ethernet fabric becomes the bottleneck—NDR InfiniBand removes that constraint entirely.
• Quantum 2 Generation Switch Fabric: Purpose-engineered for sub-microsecond latency in east-west traffic patterns. GPU-to-GPU communication over InfiniBand latency measured in hundreds of nanoseconds, not milliseconds. This directly translates to faster convergence in distributed training and reduced wall-clock time per compute job.
• Compact 1U Form Factor (7.00 x 10.00 x 14.00 inches): The MQM9790-NS2F weighs 47.76 lb and integrates high-density port capacity without requiring oversized chassis. Fits standard data center rack infrastructure alongside compute nodes, storage controllers, and monitoring appliances without special structural reinforcement.
• Industry-Standard InfiniBand Ecosystem: Compatible with any device or NIC carrying an InfiniBand port—GPUs, storage arrays, NVMe SmartNIC controllers, and compute nodes from multiple vendors. NVIDIA's MLNX_OFED driver stack and OpenFabrics standards ensure software portability across heterogeneous deployments.
• Enterprise Redundancy and Monitoring: Built-in switch management, out-of-band Ethernet port for device access, and support for InfiniBand Subnet Manager protocols. Dual-redundant power inputs allow seamless maintenance without downtime. Health monitoring via SNMP and syslog integration with NOC tools.
• Proven Supply Chain and Manufacturing: Sourced direct from the manufacturer. No grey-market, no parallel imports—factory-new units with full technical support and replacement availability through established distribution channels.

Integration & Compatibility

The MQM9790-NS2F integrates into any data center that already runs InfiniBand or is migrating to InfiniBand fabric. NVIDIA GPUs (A100, H100, H200), AMD EPYC CPUs with InfiniBand NICs, and third-party storage systems (Pure Storage, NetApp, DDN) with InfiniBand connectivity plug directly into the switch. MLNX_OFED drivers on Linux, FreeBSD, and Windows Server provide transport protocols (RDMA, RC, DC, UD) required by HPC applications, machine-learning frameworks, and distributed storage daemons. Existing InfiniBand management tools (InfiniBand Subnet Manager, Fabric Performance Analysis tools, and monitoring dashboards) work unchanged.

What's in the Box

Exact package contents are not detailed in the manufacturer specification provided. Contact presales engineering for shipping manifest details if required for procurement documentation.

Frequently Asked Questions

Q: Is the MQM9790-NS2F suitable for Ethernet-based data centers?

A: No. The MQM9790-NS2F is an InfiniBand-only fabric. It requires InfiniBand network interface cards (NICs) on all connected devices. If your infrastructure is built on Ethernet, you would need a separate InfiniBand fabric or a bridge architecture using InfiniBand-to-Ethernet gateways—both of which introduce latency penalties. Assess your existing network investment before committing.

Q: What is the maximum number of ports and devices this switch can support?

A: The Quantum 2 platform supports high port density across its product line. Consult the manufacturer datasheet or presales engineering for the exact port count on the MQM9790-NS2F model to confirm capacity for your cluster size.

Q: Does the MQM9790-NS2F include a warranty?

A: Warranty terms are determined by the distributor or manufacturer at time of purchase. Confirm coverage period with your sales contact—typical enterprise agreements include 1–5 year options with parts and support.

Q: Can this switch replace an existing Ethernet core switch?

A: Only if your cluster is InfiniBand-native or you are deliberately transitioning to InfiniBand fabric for compute interconnect. Many hybrid deployments run Ethernet for general management traffic (provisioning, monitoring, out-of-band access) and InfiniBand exclusively for compute-to-compute and compute-to-storage. Mixing architectures requires thoughtful network design.

Q: What cooling and power requirements apply?

A: Full power consumption and thermal envelope details are not provided in the available specification. Verify power draw and cooling airflow requirements with the manufacturer before integrating into a power-constrained or thermally-limited rack layout.

Q: Is the MQM9790-NS2F NDAA Section 889 compliant?

A: NDAA compliance is not stated in the available evidence. Contact the manufacturer or your distributor to confirm Section 889 compliance if NDAA requirements apply to your procurement.

Jerry Tildsen

I've built several multi-node GPU clusters over the past decade, and the jump from Ethernet to InfiniBand fabric fundamentally changes how you think about compute architecture. The MQM9790-NS2F delivers 400Gbps port throughput with sub-microsecond fabric latency—that's the spec that matters most for AI training pipelines and HPC workloads. When you're running distributed training across 32 or 64 GPUs, every millisecond of communication latency compounds across gradient synchronization and collective operations. InfiniBand cuts that penalty dramatically compared to traditional Ethernet.

Technical Highlights:

• NDR (400Gbps) Throughput per Port: Eliminates the bandwidth ceiling that forces you to overprovision compute nodes or tolerate serialized training steps. When moving 200GB+ model checkpoints or aggregate loss values from dozens of accelerators, 400Gbps means sub-second synchronization latency instead of minutes of wall-clock overhead.
• Quantum 2 Fabric Architecture with Sub-Microsecond Latency: GPU-to-GPU latency in the 300–500 nanosecond range (not milliseconds) fundamentally changes scaling efficiency. Distributed training convergence improves visibly when collective operations complete in microseconds instead of milliseconds.
• 1U Footprint, 47.76 lb Weight: Fits alongside standard compute nodes and storage controllers in rack space. No special infrastructure, no custom cabinets—integrates into existing data center racks with standard power distribution and cooling.

Deployment Considerations:

• InfiniBand is a deliberate infrastructure choice—every node must have an InfiniBand NIC. You cannot mix Ethernet and InfiniBand on the same fabric without bridge appliances, which reintroduce latency. Confirm your entire cluster ecosystem is InfiniBand-ready before deployment.
• Power draw and thermal envelope details are not available in the specification—verify with the manufacturer before committing rack space, especially if you're power-limited or running a dense installation with thermal constraints.

The MQM9790-NS2F is the right foundation for GPU-dense HPC clusters and large-scale distributed AI training environments where microsecond-level latency and 400Gbps throughput directly translate to training throughput and job completion time. It's not a general-purpose switch—use it where InfiniBand's low-latency guarantee is worth the ecosystem commitment.