NVIDIA Quantum-2 QM9700 InfiniBand Switch

The NVIDIA Quantum-2 QM9700 is a 1U 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, internally managed by an on-board subnet manager running NVIDIA MLNX-OS.

Dell sells it as part of its AI switch lineup under NVIDIA's own name - there is no Dell PowerSwitch designation for this platform. It is the 400Gb/s InfiniBand tier, and the alternative to an Ethernet fabric rather than a complement to one.

Sixty-four ports across thirty-two cages - count this correctly

This is the specification most often misread on the platform, and getting it wrong halves or doubles a cable order. The QM9700 has 32 OSFP connectors on the front panel and 64 ports of 400Gb/s. Each cage carries two NDR ports.

Port splitting takes it further: the same 1U chassis presents up to 128 ports of 200Gb/s NDR200. That density is what lets a two-level fat tree cover small-to-medium clusters without a third tier, which NVIDIA calls out for its effect on power consumption, latency and rack space rather than on port count alone.

Cabling and transceivers attach to cages, not to ports. Specify from the cage count and the splitter plan, not from the headline 64.

InfiniBand or Ethernet: how this decision is actually made

If you are speccing a GPU pod, this is the fork in the road, and it is worth being concrete about what changes.

InfiniBand is a managed fabric, not a set of independent switches. A subnet manager computes the routing tables for the whole fabric and distributes them. That centralisation is what makes adaptive routing, credit-based lossless flow control and deterministic latency possible without per-switch configuration. It also means the subnet manager is a design decision in its own right - see the next section.

In-network computing is the differentiator. The QM9700 carries the third generation of NVIDIA SHARP, which performs data aggregation and reduction inside the switch rather than at the endpoints. In a distributed training job the all-reduce collective is often the critical path, and doing that arithmetic in the fabric removes a round trip and a volume of traffic that would otherwise cross it. NVIDIA states SHARPv3 delivers AI acceleration 32x higher than its predecessor.

Topology freedom. The switch supports fat tree, Slim Fly, Dragonfly+ and multi-dimensional torus, and stays backward compatible with previous InfiniBand generations. Ethernet fabrics are effectively leaf-spine or leaf-spine-superspine; InfiniBand lets the topology follow the job's communication pattern.

Scale beyond a single fabric. With the optional router capability, clusters scale past 40,000 nodes in a single fabric - well beyond the previous generation's limit.

The counter-argument is real and should be weighed. Ethernet fabrics are built from skills your team already has, share tooling with the rest of the estate, and reach the same headline speeds - the SN5610 does 800GbE and the SN4700 does 400GbE, both with RoCE for lossless RDMA and both on the Spectrum-X platform in the SN5610's case. NVIDIA now builds seriously for both paths. The honest summary is that InfiniBand still wins on collective performance and on deterministic behaviour at scale, Ethernet wins on operational familiarity and on convergence with everything else in the data centre, and the storage and management networks are Ethernet either way. The Dell AI fabric switch guide sets the comparison out in full. Both paths sit side by side on our AI fabric switch lineup.

Internally managed - what that buys and what it costs

The QM9700 is the internally managed variant. It carries an on-board subnet manager that brings up to 2,000 nodes out of the box, running the MLNX-OS software package with full chassis management through CLI, web UI, SNMP or JSON interfaces.

For a cluster inside that 2,000-node ceiling this removes an entire component from the design: no separate fabric-manager host, no separate licence, no additional failure domain. Rack it, cable it, and the fabric comes up.

The QM9790 is the same hardware without the on-board manager - externally managed, expecting NVIDIA Unified Fabric Manager (UFM) somewhere in the fabric. Above a couple of thousand nodes, or in an estate that already runs UFM for its monitoring and preventative-maintenance features, the QM9790 is the correct choice and the on-board manager is redundant. Below that, and without an existing UFM deployment, the QM9700 is simpler. That is the whole difference between the two models.

Copper first: NVIDIA's own power argument

One line in the datasheet is worth acting on. NVIDIA notes that using copper cabling instead of optical transceivers cuts per-link power from 30 watts to 0-3 watts, and that the reduction also improves latency and reliability while lowering total cost of ownership. The recommended pattern is to group switches with short-reach copper into what NVIDIA calls a virtual modular switch, arranged as a fat tree between levels 1 and 2 or levels 2 and 3.

At 64 ports per switch, the difference between an all-optical and a copper-where-possible build is measured in kilowatts across a pod. Reach is the constraint - passive copper runs a few metres - so the physical layout of the racks determines how much of the fabric can be copper. That is a question to settle at rack-elevation time, not at cabling time.

Physical planning

The chassis is 1.70 in high, 17.00 in wide and 26.00 in deep, weighing 32.0 lbs, in a 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.

Cooling is a 6+1 hot-swappable fan unit and airflow can be front-to-rear or rear-to-front. The operating envelope depends on direction: 0 to 35 C in forward airflow, 0 to 40 C in reverse. That five-degree difference is a genuine planning input in a dense rack. Non-operating range is -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.

Where it sits in the range

The QM9790 is the externally managed twin. The Quantum-X800 Q3400-RA is the next generation - 144 ports of 800Gb/s in 4U, on 200Gb/s-per-lane SerDes, and the correct choice for a trillion-parameter-scale build. On the Ethernet side the comparable tiers are the SN4700 and SN5400 at 400GbE and the SN5610 at 800GbE. Whichever fabric you choose, the out-of-band management network is Ethernet, and that is the SN2201.

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

Specifications