NVIDIA Quantum-2 QM9700
The Quantum-2 QM9700 is a 1U, fixed-configuration NDR InfiniBand switch built around the Quantum-2 ASIC: 64 ports of 400 Gb/s NDR across 32 twin-port OSFP cages, 51.2 Tb/s of non-blocking aggregate throughput, and a 64-way radix that makes it the standard building block for H100 and H200 GPU fabrics. It ships as a managed switch (onboard x86 subnet manager, up to 2,000 nodes) or as an externally managed switch that hands subnet management to NVIDIA UFM. It supports SHARPv3 in-network reduction, adaptive routing and SHIELD self-healing link recovery. The same chassis serves as either leaf or spine in a two-tier fat-tree design, which is how most H100/H200 clusters still spec their fabric in 2026.
- In rail-optimized H100/H200 topologies, a common reference design uses 32 leaf switches (one per rail group) with a full-mesh spine layer; a two-tier, non-blocking fat tree built from 64-port QM9700 leaves and spines is commonly sized in the range of ~1,000 nodes / ~8,000 GPUs before oversubstription or a third tier is needed. Treat this as an illustrative topology size, not a vendor-guaranteed capacity number.
- Pick MQM9700 (managed) if you want the onboard subnet manager and are not already running UFM; pick MQM9790 (externally managed) if UFM is already part of the fabric management stack, it is typically the lower-cost SKU per unit.
- Budget cable and optics cost separately from the switch line item, and budget it early: OSFP DAC/AOC and transceivers routinely dominate NDR fabric spend.
- QM9700 remains the safe default for H100/H200 NDR clusters going into new builds in 2026: it is a mature, broadly stocked platform with a deep third-party optics ecosystem. Quantum-X800 is the newer step up for 800G, Blackwell-class endpoints, not a drop-in replacement.
How many GPUs can a single QM9700 fabric support?
One switch has 64 ports of 400 Gb/s NDR. Fabric-wide capacity scales with the number of switches in a leaf/spine design, not a single unit: the onboard subnet manager on a managed QM9700 tops out at 2,000 nodes, and two-tier non-blocking fat trees built from 64-port leaves and spines are commonly sized in the low thousands of GPUs before a third tier is required.
Quantum-2 QM9700 vs Quantum-X800: which one for a new build?
QM9700 runs NDR at 400 Gb/s per port and pairs naturally with H100/H200 and ConnectX-7; it is the mature, widely available choice for 2026 procurement. Quantum-X800 doubles port speed to XDR 800 Gb/s, targets Blackwell-class GPUs with ConnectX-8/9, and is the platform to specify for new 800G-endpoint builds.
Managed or unmanaged QM9700?
Managed (MQM9700-NS2F) ships with an onboard x86 subnet manager good for up to 2,000 nodes and is the simpler out-of-box option. Unmanaged/externally-managed (MQM9790-NS2F) relies on NVIDIA UFM running elsewhere and is typically the lower-priced unit, the right call once UFM is already deployed.
Is the port count 64 or 32?
Both numbers are accurate for different things. The chassis has 32 physical OSFP cages, but each cage is a twin-port connector carrying two independent 400 Gb/s lanes, giving 64 usable 400G ports, or up to 128 ports if split down to 200G.