NVIDIA Mellanox MQM8790-HS2F Quantum HDR 200Gb/s InfiniBand Switch - 40 x QSFP56 ports, 1U, externally managed, P2C airflow. The MQM8790-HS2F is the externally managed member of the NVIDIA Quantum QM8700 series (NVIDIA ordering P/N 920-9B110-00FH-0D0; confirm on the label). It provides 40 HDR 200Gb/s QSFP56 ports in 1U with 16Tb/s of non-blocking switching capacity and sub-130ns port-to-port latency. Externally managed means it ships without a management CPU and without management ports - the fabric must have a subnet manager elsewhere. The F suffix denotes P2C forward airflow. InfiniBand only. Refurbished, supplied by T.E.S IT-SOLUTIONS.
Engineering Context
The NVIDIA Quantum HDR switch device forwards 7.2 billion packets per second - 390 million pps per port - at sub-130ns port-to-port latency across 40 QSFP56 ports of 200Gb/s. Paired with ConnectX-6 adapters it also supports HDR100, using two pairs of two lanes per physical port, so one QM8790 can present up to 80 ports of 100Gb/s and serve more than 40 servers per rack without escalating to a three-level fat tree. Quantum silicon implements SHARP in-network computing for MPI and NCCL collective offload, plus adaptive routing, advanced congestion management and SHIELD self-healing. NVIDIA states that externally managed systems come without the CPU and management ports and are administered using firmware tools: the QM8790 exposes only a front micro-USB I2C port, with no Ethernet management port and no console.
Deployment and Use Cases
Cost-optimised leaf and spine positions in HDR fat-tree fabrics where a managed QM8700 or an NVIDIA UFM appliance already runs the subnet manager. Expansion of an existing HDR cluster without adding another management plane. AI training fabrics running NCCL collectives over GPUDirect RDMA for distributed PyTorch, Megatron-LM and DeepSpeed on A100-generation GPU nodes. HPC clusters running MPI (HPC-X, Open MPI, MPICH). NVMe-oF and parallel filesystem backends - Lustre, IBM Spectrum Scale, BeeGFS - at 200Gb/s per host. HDR100 double-density racks fanning out to 80 x 100Gb/s endpoints.
Technical Specifications
Brand / MPN: NVIDIA Mellanox, MQM8790-HS2F
NVIDIA ordering P/N: 920-9B110-00FH-0D0 (confirm on label)
Platform: NVIDIA Quantum, QM8790 series, 1U fixed configuration
Ports: 40 x QSFP56, HDR 200Gb/s per port
Aggregate capacity: 16Tb/s non-blocking
Port-to-port latency: sub-130ns
Packet rate: 7.2 billion packets/s (390 million pps per port)
HDR100 split: up to 80 x 100Gb/s ports with ConnectX-6
Native port speeds: 40, 56, 100, 200Gb/s
Protocol: InfiniBand only
In-network computing: SHARP, adaptive routing, congestion control, SHIELD
Management: externally managed - no CPU, no Ethernet management port, no console; front micro-USB I2C only
Subnet manager: not on board - requires OpenSM on a host, a managed switch, or NVIDIA UFM
Airflow: P2C (power-to-connector) - air enters at the PSU side, exits at the port side
Power supplies: 2 x hot-swap redundant AC, C14 inlets
Fans: 6 x hot-swap, N+1 redundant
Physical: 1U, 44 x 433 x 591mm, 12.5kg
Condition: Refurbished - tested by T.E.S IT-SOLUTIONS
Compatibility and Parent Systems
Endpoint adapters: ConnectX-6 and ConnectX-6 VPI at HDR 200Gb/s and HDR100, ConnectX-7 at HDR, BlueField-2 InfiniBand DPUs; ConnectX-5 connects at EDR 100Gb/s. Cabling is QSFP56: MCP1650-H passive DACs, MFS1S00-H AOCs, MMA1T00-HS transceivers and MCP7H50-H / MFS1S50-H HDR-to-2x HDR100 splitters. Subnet management partner: MQM8700 (internally managed, on-board SM for up to 2,048 nodes), NVIDIA UFM, or OpenSM on a Linux host with MLNX_OFED. Predecessor: Switch-IB 2 SB7800 / SB7890 at EDR. Successor: Quantum-2 QM9700 / QM9790 at NDR 400Gb/s on OSFP cages.
What This Is NOT
NOT a managed switch. No management CPU, no Ethernet management port, no console. Without a subnet manager elsewhere in the fabric the links will not come into service.
NOT the MQM8700. MQM8700-HS2F is the internally managed model with an x86 CPU, MLNX-OS and an on-board SM for up to 2,048 nodes. Same silicon, same 16Tb/s, different management model.
NOT an Ethernet or VPI switch. Quantum is InfiniBand only; the Ethernet equivalent is Spectrum-2 / Spectrum-3.
NOT an NDR switch. Per-port maximum is HDR 200Gb/s on QSFP56; NDR 400Gb/s requires Quantum-2 with OSFP cages.
NOT a C2P airflow unit. The F in -HS2F is P2C; the reverse-airflow model is MQM8790-HS2R. PSU and fan modules are airflow-specific.
NOT supplied with transceivers, cabling, rail kit or software entitlements unless listed on the quotation.
Frequently Asked Questions
Q. What does externally managed mean on the MQM8790-HS2F?
NVIDIA builds InfiniBand systems in internally and externally managed forms. Externally managed systems come without the management CPU and without management ports and are administered using firmware tools. The MQM8790-HS2F therefore has no on-board subnet manager: an SM must run elsewhere - OpenSM on a host with MLNX_OFED, a managed MQM8700, or NVIDIA UFM - or the fabric will not come up.
Q. What is the difference between MQM8790-HS2F and MQM8700-HS2F?
Silicon, port count, 16Tb/s capacity and sub-130ns latency are identical. MQM8700-HS2F adds an x86 dual-core CPU running MLNX-OS with an on-board subnet manager for up to 2,048 nodes, plus Ethernet management and console ports. MQM8790-HS2F omits all of that.
Q. How many ports and how much capacity does the MQM8790-HS2F have?
40 QSFP56 ports at HDR 200Gb/s, 16Tb/s non-blocking capacity, 7.2 billion packets per second and sub-130ns port-to-port latency. With ConnectX-6 adapters and splitter cabling it presents up to 80 ports of HDR100 100Gb/s.
Q. Is the MQM8790-HS2F port-to-connector or connector-to-power airflow?
P2C - power-to-connector forward airflow. Air enters at the power-supply side and exhausts at the cabled port side. The reverse-airflow model is MQM8790-HS2R. Fan and PSU modules are airflow-specific and every module in the chassis must match.
Q. Does the MQM8790-HS2F support SHARP in-network computing?
Yes. The NVIDIA Quantum switch device implements SHARP collective offload, executing MPI and NCCL reduction operations inside the switch silicon, alongside adaptive routing, congestion management and SHIELD self-healing.
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 Quantum switch is bench-verified before shipment: power-on, 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 subnet manager placement and airflow planning.