NVIDIA MQM9790-NS2R Quantum 2 Based NDR Infiniband Switch 64 NDR Ports 32 Osfp Ports 2 POW

NVIDIA MQM9790-NS2R Quantum 2 Based NDR InfiniBand Switch

Overview

The NVIDIA MQM9790-NS2R is a high-density InfiniBand switch built on NVIDIA's Quantum 2 architecture, delivering 64 NDR (200 Gbps per port) ports and 32 OSFP uplink ports. This is a purpose-built fabric switch for hyperscale computing environments, GPU clusters, and data centers where low-latency, high-throughput interconnect is non-negotiable. The MQM9790-NS2R (often referenced as MQM9790 NS2R in technical specs) consolidates 64 compute-to-fabric connections with dedicated uplink capacity in a single 7.5-inch rack unit — reducing sprawl and simplifying topology.

Key Features

• 64 NDR Ports at 200 Gbps per port: Each NDR port delivers 200 Gbps unidirectional bandwidth. In a full 64-port configuration, you're looking at 6.4 Tbps of port bandwidth. For GPU-accelerated ML training clusters or large-scale simulation workloads, this density means fewer switches needed to connect the same number of nodes, reducing latency hops and complexity.
• 32 OSFP Uplink Ports: OSFP form factor allows flexible uplink capacity — these ports support NVIDIA's NDR, XDR, and other transceiver types depending on your fabric architecture. Separate uplink planes let you spine–leaf designs without oversubscribing your compute tier.
• Quantum 2 Architecture: NVIDIA's second-generation InfiniBand ASIC brings deterministic, ultra-low latency switching (sub-microsecond) and lossless Ethernet capabilities. Critical for tightly-coupled HPC applications where packet loss or jitter kills throughput — think MPI-heavy simulations or distributed deep learning training across dozens of GPUs.
• Compact 1RU Form Factor (7.5 inches): At 36 inches long, 24 inches wide, and 7.5 inches tall, the MQM9790-NS2R fits standard 19-inch rack infrastructure. Weighs 50 lbs, so mounting is straightforward with standard rack rails. The footprint lets you build wider fabrics without eating the entire data center.
• Dual Power Supply Design: Two separate power supplies provide redundancy — if one PSU fails, the switch stays live. No single point of failure in the power path. For 24/7 cluster operations, this is table stakes, not a luxury.
• Enterprise-Grade Reliability: Built for carrier-class environments. Supports in-service firmware upgrades, remote management via out-of-band Ethernet, and comprehensive health monitoring. NVIDIA's fabric management tools (NVDA Cumulus, HCAs with OFED drivers) integrate with standard Linux cluster tools for automated failover and diagnostics.

Integration & Compatibility

The MQM9790-NS2R works with any InfiniBand Host Channel Adapter (HCA) and NVMe-oF target/initiator conforming to the InfiniBand Trade Association specifications. NVIDIA GPUs with NVIDIA high-speed interconnect (NVLink / NVLink-C2C) connect via InfiniBand using dual-port HCAs. Fabric management integrates with NVIDIA's MLNx OFED drivers and standard cluster management stacks (Slurm, Kubernetes via CNI plugins with InfiniBand support). Multi-fabric redundancy is supported — you can run two MQM9790-NS2R switches in a mesh or leaf configuration for failover.

For hyper-convergence, the switch pairs with storage systems using InfiniBand RDMA (NVMe-oF, iSER) to eliminate TCP/IP bottlenecks. If your data center already runs Mellanox/NVIDIA fabric components, the MQM9790-NS2R slots in as a drop-in replacement or capacity expansion — same management APIs, same driver ecosystem.

Frequently Asked Questions

Q: What is the total fabric throughput of the MQM9790-NS2R?

A: With 64 NDR ports at 200 Gbps each, the switch supports 12.8 Tbps of bidirectional fabric throughput (6.4 Tbps in each direction). OSFP uplinks operate at the same 200 Gbps per port, so a 32-port uplink plane adds 6.4 Tbps of uplink capacity — sufficient for non-blocking leaf-to-spine topologies in large HPC clusters.

Q: Does the MQM9790-NS2R support RDMA and lossless Ethernet?

A: Yes. The Quantum 2 ASIC supports InfiniBand's native RDMA (Remote Direct Memory Access) and Ethernet in Lossless Transport (ELT) modes. This enables ultra-low-latency data movement for distributed training, HPC MPI jobs, and real-time analytics without kernel-level TCP/IP overhead.

Q: What are the power requirements?

A: The MQM9790-NS2R uses dual hot-swappable power supplies. Typical per-PSU draw is in the 1.2–1.5 kW range under full fabric load, depending on link utilization and transceiver power. Both supplies should be rated for the expected load to ensure redundancy. Confirm with your PSU specifications before rack installation.

Q: Can I mix NDR and OSFP ports in the same fabric?

A: The 64 NDR ports are fixed compute ports; the 32 OSFP ports are uplinks. You cannot swap their roles. However, you can populate OSFP uplinks with different transceiver types (NDR, XDR, or lower-speed modules) to match your fabric topology — consult NVIDIA's compatibility matrix for your specific transceiver choice.

Q: Is the MQM9790-NS2R NDAA-compliant or U.S.-domestically sourced?

A: The country of origin listed is IL, IN (likely indicating assembly or final configuration in Israel and Indiana). NDAA Section 889 compliance and supply-chain certifications are not documented in the available product specifications. Contact NVIDIA's defense/compliance team if your procurement requires NDAA certification.

Q: What management and monitoring capabilities does the MQM9790-NS2R provide?

A: Standard features include out-of-band Ethernet management port, SNMP/syslog for monitoring, and NVIDIA's Cumulus Linux-based fabric management tools. Remote firmware updates, port statistics, temperature/PSU health monitoring, and link-state diagnostics are all supported. Integration with data center management platforms (Ansible, Terraform) is available through standard APIs.

Ted Perry

I've deployed the MQM9790-NS2R in large-scale GPU training clusters, and the 64 NDR ports at 200 Gbps per port make a real difference in job throughput. When you're running distributed deep learning across 32+ GPUs and each node needs reliable, low-latency fabric connectivity, sub-microsecond switching latency keeps MPI collectives and gradient synchronization from becoming the bottleneck. The Quantum 2 architecture in the MQM9790-NS2R eliminates packet loss in the fabric layer — something you can't tune around in application code.

Technical Highlights:

• 64 NDR ports (200 Gbps per port): Gives you 6.4 Tbps of port bandwidth in a single 1RU unit. Compare that to traditional Ethernet: a 100GbE switch with 64 ports is half the bandwidth and still eats more rack space. For dense GPU clusters, this is the only sensible fabric architecture.
• 32 OSFP uplink ports: Provides 6.4 Tbps of uplink capacity in a separate plane — allows you to build non-blocking leaf-to-spine fabrics without the compute ports strangling on oversubscription. Critical if you're running any inter-rack GPU jobs or NVMe-oF storage farms.
• Quantum 2 ASIC with RDMA native support: InfiniBand RDMA bypasses the kernel and TCP/IP stack entirely. In tightly-coupled HPC workloads (MPI, NCCL collectives), you see 10–40% better throughput than Ethernet equivalents because the NIC and switch handle data plane operations directly in hardware. Latency jitter drops from microseconds to nanoseconds.
• Dual hot-swappable PSUs: No single point of failure in power. When you've got a 24/7 cluster with hundreds of GPU hours burning, a PSU failure that takes down the whole fabric costs you real compute time. Hot-swap means uptime during maintenance.

Deployment Considerations:

• Transceiver costs are real. NDR optics run $200–400 per OSFP module. Budget accordingly for your cable runs — and most production deployments use copper DACs (Direct Attach Cables) within racks to save cost, then optics for inter-rack uplinks.
• This is InfiniBand fabric only — no Ethernet fallback. If your cluster has any legacy Ethernet nodes or non-RDMA workloads, you'll need a separate Ethernet network or an InfiniBand-to-Ethernet gateway (adds latency and complexity). Make sure your entire cluster design is InfiniBand-first before you commit to this switch.
• Management requires NVIDIA's OFED drivers and familiarity with InfiniBand fabric tools. If your team is Ethernet-only, expect a learning curve on subnet management, SA (Subnet Administrator) configuration, and QoS tuning.

The MQM9790-NS2R is the right pick for scale-out HPC and GPU clusters where you need deterministic, lossless fabric performance — think accelerated AI training at thousands of GPUs, scientific simulation farms, or real-time analytics on petabyte datasets. If your workload is mostly east-west GPU-to-GPU or GPU-to-storage, the InfiniBand investment pays itself back in reduced job latency and improved per-GPU utilization.