Specifications
System Model
Q3400-RA (Quantum-X800 family)
NVIDIA SKU
920-9B36F-00RX-8S0
Switch ASIC
Quantum-3
Ports
144 XDR ports at 800Gb/s, or 400Gb/s per port
Cages
72 OSFP cages (2 ports per cage)
Switching Capacity
115.2 Tb/s
SerDes
200Gb/s per lane
Form Factor
4U, 19-inch rack mount, standard depth
Dimensions
175 mm (H) × 438 mm (W) × 850 mm (D)
Weight
57.2 kg
Power Supplies
8 × hot-swappable PSU, 200-240 VAC input
Fans
10 replaceable fan modules
Airflow
Connector-to-Power (C2P)
Max Power
6,327 W with active cables fitted
Management CPU
Intel CFL 4-core i3-8100H at 3.0 GHz
Software
NVOS, with dedicated UFM in-band management port
In-Network Computing
SHARP v4, adaptive routing, telemetry congestion control
Security
CPU / CPLD / switch IC rooted in IRoT
Scale
Two-level fat tree supporting up to 10,368 ConnectX-8 NICs
Operating Temperature
0 to 35 °C, 10-85% non-condensing humidity
Altitude
Up to 3,050 m
Overview
The NVIDIA Quantum-X800 Q3400-RA is the high-radix tier of the Quantum-X800 InfiniBand platform: a 4U air-cooled switch carrying 144 XDR ports at 800Gb/s across 72 OSFP cages, for 115.2 Tb/s of aggregate switching capacity in a single chassis. It is built on the Quantum-3 ASIC and the same 200Gb/s-per-lane SerDes that feeds the ConnectX-8 SuperNIC.
A single Q3400-RA terminates a complete rail-optimised leaf layer. In a two-level fat tree it connects up to 10,368 ConnectX-8 SuperNICs with low latency and predictable job locality, which is the topology that catches most trillion-parameter training clusters. Hardware-based In-Network Computing with SHARP v4 performs reduction operations inside the switch fabric, removing that traffic from the host entirely.
Power is delivered by eight hot-swappable supplies on 200-240 VAC with connector-to-power airflow, and control runs through a dedicated OSFP in-band management port kept separate from the data ports so that full port allocation stays available for the compute fabric.
Key Benefits
Highest port density in the Quantum line. 144 XDR ports at 800Gb/s in 4U means fewer chassis and fewer optical hops per fat tree. Native SHARP v4 reduces collective-operation latency for all-reduce heavy training loops. Rail-optimised ready. The port count and airflow shape are designed for rail-aligned leaf layers against ConnectX-8 SuperNICs. Serviceable in production. Eight replaceable PSUs, ten replaceable fans, and an isolated UFM management port keep fabric administration off the data plane.
Applications
Front-end and back-end AI training fabrics for trillion-parameter models; two-level fat-tree leaf layers connecting 10,000+ ConnectX-8 SuperNICs; HPC clusters needing deterministic all-reduce via SHARP v4; bridging new XDR compute pods into existing Quantum-2 NDR storage and management infrastructure.
Request a Quote — NVIDIA QUANTUM-X800 Q3400-RA XDR INFINIBAND SWITCH — 144-PORT 800GB/S QUANTUM-3
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →