Name: NVIDIA Quantum-X800 Q3400-RA / Q3200-RA InfiniBand Switch
Brand: NVIDIA

NVIDIA

NVIDIA Quantum-X800 Q3400-RA / Q3200-RA InfiniBand Switch

800Gb/s XDR InfiniBand switch family. Q3400-RA 144 ports / 4U, Q3200-RA 72 ports / 2U; SHARPv4, adaptive routing and telemetry-based congestion control.

Q3400-RA: 144 x 800G XDR ports over 72 OSFP cages, 115.2 Tb/s
Q3200-RA: 72 x 800G XDR ports over 36 OSFP cages, 57.6 Tb/s
200Gb/s SerDes — double the per-lane rate
SHARPv4 in-network collective computing

Models in this family

Q3400-RA Q3200-RA

Get a Quote +90 212 483 70 33

Authorised dealer supply · manufacturer warranty · price and lead time per configuration

Quantum-X800 is NVIDIA’s 800Gb/s XDR generation. The difference is not only speed: with 200Gb/s SerDes per lane, per-port bandwidth doubles while the physical port count stays the same. In practice that means twice the capacity at the same cabling density.

The family has two air-cooled models. Q3400-RA delivers 144 ports of 800G over 72 OSFP cages in a 4U chassis — a radix high enough to gather more than 10,000 network adapters into a single fabric with a two-level fat-tree, removing the need for a third topology level. Q3200-RA brings the same generation down to 72 ports in a 2U chassis, suited to smaller installations and to the leaf layer.

The family also includes busbar-powered (Q3401-RD) and co-packaged-optics (Q3450-LD) variants, intended for installations whose rack and power architecture is designed around them. Which variant fits is assessed together with rack power distribution and cooling capacity.

One point deserves attention when planning power: the difference between passive and active cabling. The Q3400-RA draws a typical 2,900 W with passive copper, but reaches a maximum of 7,000 W in a fully active-optical configuration. Do not finalise the rack power budget before the cable type is settled.

Specifications

Specification	Q3400-RA	Q3200-RA
Connectivity and capacity
Ports	144 x XDR 800G (72 OSFP cages)	72 x XDR 800G (36 OSFP cages)
Switching capacity	115.2 Tb/s	57.6 Tb/s
Chassis	4U, 19" rack	2U, 19" rack
Connector	OSFP	OSFP
Fabric capabilities
SHARPv4 (in-network collectives)	Yes	Yes
Adaptive routing	Yes	Yes
Telemetry-based congestion control	Yes	Yes
Performance isolation	Yes	Yes
Dedicated UFM port	Yes	Yes
Physical and electrical
Dimensions (H x W x D)	177.8 x 438 x 850 mm	88 x 438 x 680 mm
Weight	60 kg (with 8 PSUs)	49 kg
Power supplies	8, replaceable	4, replaceable
Fans	10	5
Input voltage	200-240 VAC, 10 A, 50/60 Hz (per PSU)	200-240 VAC, 10 A, 50/60 Hz (per PSU)
Typical power (passive cables)	2,900 W	862 W
Max power (active cables)	7,000 W	3,486 W
Airflow	Reverse	Reverse
Operating temperature	0-35 °C	0-35 °C

Specifications follow the manufacturer datasheet. The exact configuration is confirmed in your quote.

Connectivity Solutions

XDR spine and leaf layers

The 144-port radix of the Q3400-RA is enough to gather more than 10,000 network adapters into a single fabric with a two-level fat-tree. Because no third topology level is needed, both latency and the number of spine switches drop. The Q3200-RA brings the same generation to smaller installations and to the leaf layer.

Goes with InfiniBand adapters

Optics and cabling

At the XDR tier links are built over OSFP cages; passive copper DAC is used at short reach and active optics between racks. Active cables raise switch power draw substantially — finalise the power and cooling plan only after the cable type is decided.

Goes with InfiniBand transceivers DAC and AOC cables

Management and telemetry

The system runs NVOS and provides a dedicated management port for UFM. Fabric telemetry feeds the congestion-control decisions, reducing the variance in job completion times under heavy collective traffic.

Goes with HPC Network Design

Features

Double the lane rate with 200Gb/s SerDes

Quantum-X800 signals at 200Gb/s per lane. That means twice the bandwidth at the same port count: per-port throughput reaches 800Gb/s without increasing the number of cables and optics.

SHARPv4

In-network aggregation of collective MPI operations moves to its fourth generation. For all-reduce heavy distributed training, the benefit grows with the node count.

Telemetry-based congestion control

Congestion is managed from live fabric telemetry rather than fixed thresholds. Performance isolation keeps jobs sharing the same fabric from disturbing each other's completion times.

High radix, shallower topology

144 ports reach more than ten thousand endpoints in a two-level fat-tree. Dropping the third level cuts both end-to-end latency and the total cost of switches and optics.