GS-NET-02Engineering datasheet

NVIDIA Quantum-X800

Statusshipping
Verified2026-07-10

Quantum-X800 is NVIDIA's XDR-generation InfiniBand switch platform, built around the Quantum-3 ASIC and paired with Blackwell-class GPUs (GB200, B200, B300) over ConnectX-8/9 SuperNICs running at up to 1.6 Tb/s per GPU. The Q3400-RA is the 4U flagship: 144 ports of 800 Gb/s XDR across 72 twin-port OSFP cages, 115.2 Tb/s of non-blocking aggregate throughput, and sub-100 ns port-to-port latency. A smaller 2U Q3200-RA (72 ports, 57.6 Tb/s) covers lower-radix leaf roles or bridges legacy Quantum-2 fabrics into an XDR build. Both chassis run as leaf or spine in a two-tier fat tree, and NVIDIA's own reference design scales a single fat tree built from Q3400-RA switches to 10,368 NICs.

144 × 800 Gb/s XDR
Port configuration · Q3400-RA, 72 twin-port OSFP cages; Q3200-RA: 72 × 800 Gb/s over 36 cages, 2U
115.2 Tb/s
Aggregate throughput · Q3400-RA, non-blocking; Q3200-RA: 57.6 Tb/s
< 100 ns
Port-to-port latency · reported
2,900 W typical
Power · Q3400-RA, passive cables; up to 7,000 W with active cables
4U
Form factor · Q3400-RA; Q3200-RA is 2U, both air-cooled, 19 in rack
10,368 NICs
Fabric scale · two-level fat tree reference design, ConnectX-8
Switching01/05
ASICNVIDIA Quantum-3
Switching capacity, Q3400-RA115.2 Tb/s
Switching capacity, Q3200-RA57.6 Tb/s
Radix144 ports (Q3400-RA) / 72 ports (Q3200-RA) · at 800 Gb/s XDR
SerDes200 Gb/s per lane
In-network computingSHARPv4 · 14.4 TFLOPS; 9x more in-network compute than the prior NDR platform
RoutingAdaptive routing, telemetry-based congestion control, performance isolation
Ports & cabling02/05
Connectors, Q3400-RA72 × OSFP · twin-port cages, each carrying 2 × 800 Gb/s
Connectors, Q3200-RA36 × OSFP
Management port1 dedicated OSFP · in-band UFM management, separate from data ports
Front I/OUSB 2.0 Type-A, 2 × MGT, console port
Endpoint NICsConnectX-8 / ConnectX-9 SuperNICs
Power & thermal03/05
Power, Q3400-RA2,900 W typical / 7,000 W max · typical with passive cables, max with active cables
Power, Q3200-RA862 W typical / 3,486 W max
Input200-240 VAC
Power supplies8 (Q3400-RA) / 4 (Q3200-RA) · hot-swappable
Operating temperature0-35°C · reverse (C2P) airflow
Humidity10%-85% non-condensing
Altitude3,050 m max
Acoustic, Q3400-RA78.4 dBA · at room temperature
Physical04/05
Dimensions, Q3400-RA (H x W x D)177.8 mm × 438 mm × 850 mm · 4U, 7 in × 17.2 in × 33.5 in
Weight, Q3400-RA60 kg · with 8 PSUs
Weight, Q3200-RA24 kg · with 4 PSUs
Rack mount19 in standard rack, standard depth
AirflowConnector-to-power (C2P)
Management & features05/05
SoftwareNVOS
CPUIntel Coffee Lake i3-8100H, 4 cores, 3 GHz
Fabric managementNVIDIA UFM
SecurityCPU / CPLD / switch IC hardware root of trust (IRoT)
Topologies supportedTwo-level fat tree (up to 10,368 NICs per reference design)
Field notes
  • Quantum-X800 is the fabric to specify for Blackwell/GB200-class 800G endpoints; it is not a drop-in upgrade path for existing NDR fabrics, since cabling and optics are a different generation end to end.
  • A documented GTC 2024 reference design builds a 576-GPU rail group as a 'virtual modular switch' from 4 Quantum-X800 switches in a 3-tier topology; treat single-switch radix numbers as a building block for such designs, not a cluster-sizing shortcut on their own.
  • The dedicated management OSFP port keeps UFM traffic off the data plane; budget a cable and a port for it even though it does not show up in the headline 144-port count.
  • Chassis power draw (2,900 W typical, up to 7,000 W with active cabling on the Q3400-RA) approaches the power budget of a full GPU compute node; rack-level power planning should treat the switch as a first-class consumer, not an afterthought.
Questions we get on this part

How many GPUs can a Quantum-X800 fabric support?

NVIDIA's published two-level, non-blocking fat-tree reference design for the Q3400-RA connects up to 10,368 ConnectX-8 NICs. Actual GPU count depends on topology tier count and oversubscription choices; larger builds add a third tier.

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.

Q3200-RA vs Q3400-RA, which chassis?

Q3400-RA is the 4U, 144-port spine-class chassis with 115.2 Tb/s of switching capacity. Q3200-RA is the 2U, 72-port chassis (57.6 Tb/s) built for smaller deployments or for bridging legacy Quantum/Quantum-2 infrastructure into an XDR fabric.

Is Quantum-X800 backward compatible with NDR gear?

Not at the port level. XDR runs 800 Gb/s per port over new OSFP optics and 200 Gb/s-per-lane SerDes, so it requires its own cabling and ConnectX-8/9 endpoints rather than reusing NDR transceivers or NICs.