NVIDIA Quantum-2 QM9790 InfiniBand Switch

The NVIDIA Quantum-2 QM9790 is a 1U externally managed InfiniBand switch presenting 64 non-blocking ports of 400Gb/s NDR across 32 OSFP cages, with 51.2 Tbps of aggregate bidirectional throughput and a packet rate above 66.5 billion packets per second. It carries no on-board subnet manager and expects NVIDIA Unified Fabric Manager (UFM) elsewhere in the fabric.

Hardware-for-hardware it is identical to the QM9700. The management model is the entire difference, and it is a design decision rather than a feature tier.

Externally managed: what that means and when it is the right answer

"Externally managed" is not "unmanaged". In Ethernet switching the second term means an appliance with no control plane at all; here it means the opposite - a fully featured InfiniBand switch whose routing is computed by a subnet manager running somewhere else in the fabric.

Every InfiniBand fabric has a subnet manager. It discovers the topology, computes forwarding tables for the whole subnet and distributes them, and it is what makes InfiniBand's adaptive routing and deterministic latency possible without configuring switches individually. The only question is where it runs.

On the QM9700 it runs on the switch, and handles up to 2,000 nodes. On the QM9790 it does not, and you supply it. NVIDIA's intended answer is UFM, which does considerably more than compute routes: it provisions, monitors, preventatively troubleshoots and maintains the fabric, with the stated aim of higher utilisation and lower operational cost.

Three situations make the QM9790 the correct choice:

• The cluster exceeds the on-board manager's ceiling. Past a couple of thousand nodes the on-board subnet manager stops being the answer, and paying for it on every switch in a large fabric is paying for something you will not use.
• UFM is already deployed. If the estate runs UFM, an on-board manager is a redundant component and a second thing to keep patched.
• The fabric spans multiple switches and you want one control point. Centralised provisioning and monitoring across the whole subnet is easier to reason about than per-switch managers with one designated master.

Conversely, for a single-rack or small multi-rack cluster with no existing fabric-management practice, the QM9700 removes a host, a licence and a failure domain from the design. Neither model is better; they answer different questions.

Sixty-four ports across thirty-two cages

The QM9790 has 32 OSFP connectors and 64 ports of 400Gb/s NDR - two ports per cage. Port splitting presents up to 128 ports of 200Gb/s NDR200 from the same 1U chassis.

Cables and transceivers attach to cages, not to ports. Specify from the cage count and the splitter plan. This is the most commonly miscounted specification on the Quantum-2 platform and the error is expensive in both directions.

What the fabric does that Ethernet does not

The case for InfiniBand in a GPU pod rests on three things, and all three are present on this switch.

In-network computing. Third-generation NVIDIA SHARP performs data aggregation and reduction inside the fabric rather than at the endpoints. In distributed training the all-reduce collective is frequently the critical path, and executing it in the network removes a round trip and a volume of traffic that would otherwise have to cross it. NVIDIA states SHARPv3 delivers AI acceleration 32x higher than its predecessor.

Fabric-level behaviour. Self-healing network capability, quality of service, enhanced virtual lane mapping and advanced congestion control - all computed centrally and applied consistently, which is what produces the deterministic latency that makes a synchronised workload predictable.

Topology and scale freedom. Fat tree, Slim Fly, Dragonfly+, multi-dimensional torus and others, with backward compatibility to previous InfiniBand generations. With the optional router capability, a single fabric scales past 40,000 nodes.

Against that, an Ethernet fabric is built from skills the team already has and converges with the rest of the data centre - the SN4700 and SN5610 cover the same speed tiers with RoCE for lossless RDMA. The storage and out-of-band networks are Ethernet regardless. The Dell AI fabric switch guide works the comparison through properly. The Quantum and Spectrum switches are listed alongside each other on our AI networking switch lineup.

Copper before optics

NVIDIA's own datasheet makes the argument: copper cabling instead of optical transceivers cuts per-link power from 30 watts to 0-3 watts, and the reduction improves latency and reliability as well as cost. The recommended pattern is to group QM9700-series switches with short-reach copper into a virtual modular switch, arranged as a fat tree between levels 1 and 2 or levels 2 and 3.

Across a pod of 64-port switches that is a kilowatt-scale difference. Passive copper reach is a few metres, so how much of the fabric can be copper is decided by the rack elevation - settle it before the cable order, not after.

Physical planning

The chassis is 1.70 in high, 17.00 in wide and 26.00 in deep, weighing 32.0 lbs, 1U rack mount. Power is two hot-swappable supplies in 1+1 redundancy at 200-240V AC, 80 PLUS Gold+ and ENERGY STAR certified; NVIDIA notes US installations need two phases totalling at least 208V. This model takes the same AC supply as the QM9700 - the DC bus bar variant in the series is the QM9701, which is a different product.

Cooling is a 6+1 hot-swappable fan unit, front-to-rear or rear-to-front. The operating envelope follows the direction: 0 to 35 C forward, 0 to 40 C reverse. Non-operating -40 to 70 C, humidity 10-85% non-condensing operating and 10-90% non-operating, altitude to 3,050 m.

Control plane is an x86 Coffee Lake i3 with 8GB of DDR4 SO-DIMM at 2,666 MT/s and a 16GB M.2 SATA SSD. Management ports are 1x USB 3.0, 1x USB for I2C and three RJ45, one of them UART - these are for chassis access, not for fabric management, which happens through UFM.

Where it sits in the range

The QM9700 is the internally managed twin and the better choice below roughly 2,000 nodes without an existing UFM deployment. The Quantum-X800 Q3400-RA is the next generation at 144 ports of 800Gb/s in 4U. On the Ethernet side, the SN4700 and SN5400 are the 400GbE tiers and the SN5610 the 800GbE. The out-of-band management network is Ethernet whichever fabric you build, and that is the SN2201.

Every unit is professionally reconditioned, fully tested and shipped with warranty. This model is quote-only. NDR-generation InfiniBand barely circulates on the secondary market, so tell us the node count, topology, subnet-manager plan and quantity and a specialist will come back to you on sourcing and availability.

Specifications