NVIDIA Quantum-2 QM9700 / QM9790 InfiniBand Switch
A 1U InfiniBand switch family with 64 x 400G NDR ports. QM9700 managed, QM9790 unmanaged; SHARP and adaptive routing supported.
NVIDIA Quantum-2 QM9700 / QM9790 InfiniBand Switch
A 1U InfiniBand switch family with 64 x 400G NDR ports. QM9700 managed, QM9790 unmanaged; SHARP and adaptive routing supported.
- 64 x 400Gb/s NDR ports over 32 OSFP connectors
- 25.6 Tb/s switching capacity in a 1U chassis
- SHARPv3 aggregates collective MPI operations inside the network
- Redundant hot-swap power supplies and fan modules
The QM9700 and QM9790 are 1U InfiniBand switches built on the NVIDIA Quantum-2 silicon. Both offer 64 links at 400Gb/s NDR over 32 OSFP connectors; the difference is on the management side. The QM9700 ships with an onboard subnet manager and can run the fabric itself, while the QM9790 leaves management to an external subnet manager.
In practice most deployments use both. One or two QM9700 units take on fabric management and QM9790 units fill out the remaining capacity — a structure that balances cost against management complexity.
Specifications
| Specification | QM9700 MQM9700-NS2F | QM9790 MQM9790-NS2F |
|---|---|---|
| Connectivity and capacity | ||
| Ports | 64 x NDR 400G (32 OSFP connectors) | 64 x NDR 400G (32 OSFP connectors) |
| Switching capacity | 25.6 Tb/s | 25.6 Tb/s |
| Management | Managed — inband / outband | Unmanaged — inband |
| Fabric capabilities | ||
| Subnet manager (onboard) | Yes | No |
| Adaptive routing | Yes | Yes |
| SHARP | Yes | Yes |
| VL2VL | Yes | Yes |
| QoS | Yes | Yes |
| Physical and electrical | ||
| Dimensions (H x W x D) | 43.6 x 438 x 660 mm (1U) | 43.6 x 438 x 660 mm (1U) |
| Power supplies | 2 (1+1 redundant), hot-swap, AC | 2 (1+1 redundant), hot-swap, AC |
| Fans | 7 (6+1 redundant), hot-swap, back-to-front | 7 (6+1 redundant), hot-swap, back-to-front |
| Max. power consumption | 1703 W | 1593 W |
Specifications follow the manufacturer datasheet. The exact configuration is confirmed in your quote.
Connectivity Solutions
Spine layer — splitting the 800G port in two
With twin-port modules, each OSFP connector carries two 400G NDR links. That turns 32 physical ports into 64 logical connections and lowers both the switch count and the rack space needed for the same spine capacity. The breakout plan is applied together with port-splitting configuration during commissioning.
From leaf layer to compute node
On links from the leaf switch to the ConnectX-7 adapter in the server, the switch side takes a finned module and the adapter side a flat-top one. Short runs use copper breakout cables; longer runs use optical modules plus trunk cabling.
Management network and subnet manager
The QM9700 manages the fabric through its onboard subnet manager, with no separate management server required. If you choose the QM9790, the subnet manager runs on a server inside the cluster, preferably redundantly. In mixed deployments, making the first switch managed is a common and economical approach.
Features
SHARPv3 — in-network computing
Part of collective MPI operations such as all-reduce is aggregated on the switch; instead of circulating between every node, the data is reduced inside the network. The gain grows with node count and is the single feature that makes the biggest difference in distributed training.
Adaptive routing
Traffic is spread across alternative paths according to live load rather than a static route. In a fat-tree topology this prevents specific links from congesting and reduces variance in job completion times.
Redundant power and cooling
1+1 redundant hot-swap power supplies and 6+1 redundant fan modules keep the switch running through a single component failure. Back-to-front airflow suits hot-aisle / cold-aisle layouts.
High port density in 1U
Fitting 64 x 400G links into a single rack unit reduces both rack space and cabling complexity at the spine layer. That density raises the importance of cable management and airflow planning.
Resources
- QM9700 product page and specifications Datasheet
- QM9790 product page and specifications Datasheet
- NVIDIA Quantum-2 InfiniBand platform Technical note
Documents open on the manufacturer's or supplier's own site.
Request a quote for this product
Tell us the quantity and the cluster it will join; we send a compatible part list with lead times — usually the same business day.
- Authorised dealer supply, manufacturer warranty
- Compatibility check for switch, adapter and optics
- Lead time per line item
Let's Build Future Together.
