English 
| Model Number: | |
|---|---|
| Switch Architecture: | |
| Parameter Specification Model Number MQM9790-NS2R Switch Architecture NVIDIA Quantum-2 Network Type: | |
| Port Configuration: | |
| Maximum Speed: | |
| Total Switching Capacity: | |
| Management Interface: | |
| Quantity: | |
MQM9790-NS2R
Mellanox
The NVIDIA Mellanox MQM9790-NS2R is a 1U rack-mount NDR InfiniBand switch based on the NVIDIA Quantum-2 switch ASIC. This model supports 32 OSFP ports with line-rate 400Gb/s per port connectivity, designed for GPU cluster interconnects in AI training facilities and high-performance computing environments. The switch operates without an internal CPU, delivering cut-through switching for reduced port-to-port latency in large-scale fabric deployments. Commercial procurement teams specify this model for top-of-rack and spine deployments where deterministic network performance is required for distributed workloads.
Parameter | Value |
Model Number | MQM9790-NS2R |
Part Number | 920-9B210-00RN-0D0 |
Switch Architecture | NVIDIA Quantum-2 IC |
Network Type | InfiniBand NDR |
Port Configuration | 32 × OSFP cages |
Port Speed Support | 40, 56, 100, 200, 400 Gb/s per port |
Maximum Port Configuration (with split) | 64 × NDR200 ports |
Switching Capacity | 25.6 Tbps |
Aggregate Bandwidth | 51.2 Tbps bidirectional |
Packet Forwarding Rate | 665 Bpps |
Host Interface | 1 × Ethernet management port |
Form Factor | 1U rack-mount |
Dimensions (W × H × D) | 438 mm × 43.6 mm × 660 mm |
Weight (with 2 PSUs) | 14.8 kg |
Typical Power Consumption (passive cables) | 640 W |
Maximum Power Consumption (active cables) | 1,610 W |
Input Voltage | 200-240 VAC, 50/60 Hz |
Operating Temperature | 0°C to 35°C (forward airflow) |
Operating Humidity | 10%-85% non-condensing |
Safety Certifications | CB, cTUVus, CE, CU, S-Mark |
EMC Certifications | CE, FCC, VCCI, ICES, RCM, CQC, BSMI, KCC |
Environmental Compliance | RoHS compliant |
The switch implements NVIDIA SHARP technology for in-network computing, reducing the load on endpoint CPUs during collective operations. Adaptive routing dynamically adjusts traffic paths to maintain consistent throughput during congestion events, which supports mixed-workload environments in enterprise data centers.
Each OSFP port supports port splitting into two NDR200 interfaces, allowing deployment of up to 64 ports of 200Gb/s connectivity without additional hardware investment. This configuration provides flexibility for connecting different server generations within the same cluster.
The MQM9790-NS2R model supports port-to-power (C2P) airflow direction, matching hot-aisle/cold-aisle containment configurations in most modern data centers. Redundant, hot-swappable power supplies and fan modules provide continuous operation during component replacement.
This switch provides GPU-direct communication capabilities for distributed training workloads, supporting direct memory access between GPUs across multiple server nodes. The 400Gb/s per-port bandwidth matches the connectivity requirements of current-generation NVIDIA H100 and A100 GPU servers.
For HPC clusters running simulation and modeling workloads, the switch delivers consistent latency characteristics required for tightly coupled parallel applications. The InfiniBand architecture supports RDMA operations for efficient inter-node communication.
The switch supports high-throughput connectivity for parallel storage systems, enabling concurrent access from multiple compute nodes in large-scale storage deployments.
The MQM9790-NS2R operates with all NVIDIA ConnectX-7 InfiniBand adapters and is backward compatible with HDR 200Gb/s and EDR 100Gb/s InfiniBand equipment. The switch works with passive copper cables, active optical cables, and OSFP transceivers from NVIDIA-approved vendors. Management integration works with standard data center monitoring systems via SNMP and the NVIDIA UFM platform.
The OSFP ports support passive direct attach copper cables up to 2 meters, active optical cables up to 100 meters, and fiber transceivers with MPO-12 connectors for multimode and single-mode fiber deployments.
The MQM9790-NS2R is an unmanaged switch. Fabric management is handled through the NVIDIA Unified Fabric Manager (UFM) platform running on a separate server, which provides visibility and configuration for the entire InfiniBand fabric.
The switch provides cut-through switching with a port-to-port latency of approximately 110 nanoseconds for 400Gb/s ports under standard operating conditions.
Yes, all ports can be configured for NDR200 operation, and port splitting can be enabled on a per-port basis to support 200Gb/s server connections where 400Gb/s is not required.
