NVIDIA Mellanox MQM9700-NS2F Quantum-2 NDR 400Gb/s InfiniBand Switch - 64 ports, 32 OSFP cages, managed, P2C airflow. The MQM9700-NS2F is the internally managed QM9700 variant of the NVIDIA Quantum-2 platform (NVIDIA ordering P/N 920-9B210-00FN-0M0; confirm the build suffix on the label). It delivers 64 NDR 400Gb/s InfiniBand ports from 32 physical OSFP connectors in a 1U chassis, with 51.2Tb/s aggregate bi-directional throughput and more than 66.5 billion packets per second. The F in -NS2F denotes power-to-connector (P2C) forward airflow: air enters at the power-supply side and exhausts at the port side. InfiniBand only. Refurbished, supplied by T.E.S IT-SOLUTIONS.
Engineering Context
Quantum-2 uses 100Gb/s-class PAM4 serdes. Each OSFP cage carries eight electrical lanes and is presented as two independent 400Gb/s NDR ports, which is why 32 cages yield 64 NDR ports; NVIDIA port-split divides them further into up to 128 x 200Gb/s NDR200 links, the densest NDR200 radix available in 1U. Native per-port rates are 40, 56, 100, 200 and 400Gb/s. QM9700 introduces SHARPv3, supporting up to 64 parallel aggregation flows - 32x the previous Quantum generation - so MPI and NCCL collective reductions execute inside the switch instead of repeatedly crossing the fabric. Adaptive routing, congestion control, QoS and SHIELD self-healing are implemented in silicon; fat tree, DragonFly+, SlimFly and Torus topologies are supported. NVIDIA publishes 51.2Tb/s and 66.5 BPPS but no port-to-port latency figure in the QM9700 datasheet; the previous HDR generation (QM8700 / QM8790) is documented at sub-130ns.
Deployment and Use Cases
Leaf and spine positions in NDR fabrics for large-scale AI training - NCCL allreduce and allgather over GPUDirect RDMA, distributed PyTorch, Megatron-LM and DeepSpeed across H100, H200 and GH200-class nodes. Compute and storage fabric in DGX H100-generation SuperPOD-style designs. HPC clusters running MPI (HPC-X, Open MPI, MPICH). NVMe-oF and parallel filesystem backends - Lustre, IBM Spectrum Scale, BeeGFS, WEKA, DDN. NDR200 breakout serving 128 x 200Gb/s endpoints from one rack unit. P2C is the correct choice where the cold aisle faces the power-supply side.
Technical Specifications
Brand / MPN: NVIDIA Mellanox, MQM9700-NS2F
NVIDIA ordering P/N: 920-9B210-00FN-0M0 (suffix varies by build)
Platform: NVIDIA Quantum-2, QM9700 series, 1U fixed-configuration switch
Ports: 64 x NDR 400Gb/s on 32 OSFP cages (2 x NDR per cage)
Port split: up to 128 x 200Gb/s NDR200
Aggregate throughput: 51.2Tb/s bi-directional; 66.5+ billion packets/s
Native port speeds: 40, 56, 100, 200, 400Gb/s
Protocol: InfiniBand NDR only
In-network computing: SHARPv3, up to 64 parallel aggregation flows
Management: internally managed - x86 Coffee Lake i3, 8GB DDR4-2666, 16GB M.2 SSD, MLNX-OS
Subnet manager: on-board SM, bring-up for up to 2,000 nodes; CLI, WebUI, SNMP, JSON
Airflow: P2C (power-to-connector) forward - air enters at the PSU side, exits at the port side
Power supplies: 2 x hot-swap 2,000W AC, 1+1 redundant, 200-240VAC, C16 (P2C PSU MTEF-PSF-AC-J)
Fans: 7 x hot-swap, N+1 redundant (P2C fan MTEF-FANF-L)
System power: 1,084W max passive cabling; 1,720W max with optics at 18.5W/port
Physical: 1U, 44 x 438 x 660mm, approx. 14.5kg; 0 C to 40 C
Condition: Refurbished - tested by T.E.S IT-SOLUTIONS
Compatibility and Parent Systems
Endpoints: ConnectX-7 (NDR OSFP and QSFP112) and BlueField-3 DPUs at 400Gb/s; ConnectX-6 at HDR 200Gb/s and ConnectX-5 at EDR 100Gb/s via splitter cabling. Media is the NVIDIA LinkX NDR OSFP range: MCP4Y10-N passive DACs, MCA4J80-N and MCA7J60-N active copper cables and 800G-to-2x400G splitters, MMA4Z00-NS multimode and MMS4X00-NS single-mode transceivers, plus OSFP-to-2x QSFP56 splitters for HDR. Air-cooled Quantum-2 switch cages take finned-top OSFP modules while ConnectX-7 cages take flat-top modules - not interchangeable. Companions: MQM9790 (externally managed), QM8700 / QM8790 at HDR, NVIDIA UFM. Predecessor: Quantum QM8700. Successor: Quantum-3 Q3200-RA / Q3400-RA at XDR 800Gb/s.
What This Is NOT
NOT an Ethernet switch. InfiniBand only; the Ethernet equivalent of this generation is Spectrum-4 (SN5600).
NOT the externally managed MQM9790. The NS2F carries the x86 CPU and on-board subnet manager; MQM9790 has neither and needs an external SM or UFM.
NOT a C2P airflow unit. -NS2F is P2C (forward); the reverse-airflow model is MQM9700-NS2R. PSU and fan modules are airflow-specific.
NOT a QSFP56, QSFP-DD or QSFP112 switch. All 32 cages are OSFP; HDR gear attaches only via OSFP-to-QSFP56 splitters.
NOT compatible with flat-top OSFP modules. Air-cooled switch ports need finned-top OSFP transceivers.
NOT 2U, and NOT supplied with OSFP transceivers, cabling, rail kit or software entitlements unless listed on the quotation.
Frequently Asked Questions
Q. How does the MQM9700-NS2F have 64 ports on 32 connectors?
Each OSFP cage carries eight 100Gb/s-class PAM4 lanes and is presented to the fabric as two independent 400Gb/s NDR ports, so 32 cages equal 64 NDR ports. NVIDIA port-split can divide them further into up to 128 x 200Gb/s NDR200 links.
Q. Is the MQM9700-NS2F port-to-connector or connector-to-power airflow?
P2C - power-to-connector, forward airflow. Cool air enters at the power-supply side and hot air exhausts at the cabled port side. The reverse-airflow model is MQM9700-NS2R. Fan and PSU modules are airflow-specific; mixing directions causes recirculation and thermal shutdown.
Q. Is the MQM9700-NS2F managed, and how many nodes can its subnet manager handle?
It is the internally managed variant, with an x86 Coffee Lake i3 running MLNX-OS and an on-board subnet manager rated for out-of-the-box bring-up of up to 2,000 nodes. Beyond that, or where central provisioning and telemetry are required, NVIDIA UFM is used instead.
Q. Which transceivers and cables does the MQM9700-NS2F take?
NVIDIA LinkX NDR OSFP media: MCP4Y10-N passive DACs, MCA4J80-N and MCA7J60-N active copper cables and splitters, MMA4Z00-NS multimode and MMS4X00-NS single-mode OSFP transceivers. Switch cages require finned-top modules; flat-top versions are for ConnectX-7 and BlueField-3.
Q. Does the MQM9700-NS2F support SHARP in-network computing?
Yes. Quantum-2 implements SHARPv3 with up to 64 parallel aggregation flows, 32 times the aggregation capacity of the previous Quantum HDR generation, offloading MPI and NCCL collective reductions into switch silicon.
Why T.E.S IT-SOLUTIONS
T.E.S IT-SOLUTIONS (tes-itsolutions.com) supplies refurbished NVIDIA Mellanox InfiniBand switching to B2B, HPC and AI customers across the EU and worldwide. Every switch is bench-verified before shipment: power-on, management response, port link verification where test topology allows, PSU and fan redundancy check and cosmetic inspection. EU stock, 1-3 business day handling and technical consultation on NDR fabric design and cable selection.