NVIDIA Quantum-X800 Q3400-RA InfiniBand Switch

The NVIDIA Quantum-X800 Q3400-RA is a 4U air-cooled InfiniBand switch presenting 144 non-blocking ports of 800Gb/s across 72 OSFP cages, with 115.2 Tbps of throughput, and it is the first NVIDIA switch generation to use 200Gb/s-per-lane SerDes. It runs NVIDIA Networking OS Software (NVOS) and carries a dedicated 400Gb/s in-band management port for Unified Fabric Manager.

This is the switch built for trillion-parameter-scale generative AI. A two-tier Quantum-X800 fat tree supports more than 10,000 host connections at 800Gb/s, and NVIDIA puts the exact figure at up to 10,368 network interface cards with minimal latency and optimal job locality.

144 ports across 72 cages - and what that does to a cable order

Each of the 72 OSFP cages carries two 800Gb/s ports. The headline 144 is the port count; the 72 is what you physically plug into.

At this scale that distinction stops being pedantry. Seventy-two cages of 800Gb/s is a cabling project in its own right, and on a chassis this size the interconnect is frequently a larger line item than most switches cost. Get the cage count, the splitter plan and the copper-versus-optics split settled before anything else, because they determine the rack elevation as much as the elevation determines them.

The management port is separate and does not consume a data port. NVIDIA separates it on the front panel specifically so the full set of standard ports stays available for the fabric, which simplifies port allocation and topology design. Note that it is itself an OSFP 400Gb/s InfiniBand port, not an RJ45 - the management plane needs InfiniBand cabling too.

The generational step: 200G SerDes, and what it unlocks

Quantum-X800 is the first NVIDIA switch platform on 200Gb/s-per-lane serialiser/deserialiser technology. That is the change underneath everything else: doubling the per-lane rate is what produces 800Gb/s per port from the same eight-lane cage, and what makes a 144-port 800Gb/s switch physically possible in 4U.

NVIDIA's stated result is 2x faster speeds and 5x higher scalability for AI compute fabrics than the previous generation. Concretely, against the Quantum-2 QM9700's 64 ports of 400Gb/s in 1U, this is 144 ports at double the rate - roughly four and a half times the switch capacity in four times the height, with the radix concentrated in one device so a two-level fat tree covers what would otherwise need three tiers.

Fewer tiers is not a packaging convenience. Every tier a packet crosses adds latency and adds a place for congestion to form, and in a synchronised collective the slowest path sets the pace for every participant.

Fourth-generation SHARP: computing inside the fabric

The Q3400-RA carries the fourth generation of NVIDIA SHARP, and this is the clearest single argument for InfiniBand on a large training cluster. SHARP performs collective operations - the aggregations and reductions that dominate distributed training - inside the network rather than at the endpoints. NVIDIA states the offload boosts performance by up to 9x.

The fourth generation adds two things that matter for current model architectures: FP8 precision, so the in-network arithmetic matches the precision the training run actually uses rather than forcing a conversion, and new collectives in ReduceScatter and ScatterGather, which are the primitives behind sharded optimiser states and tensor-parallel layouts.

Alongside it sit enhanced adaptive routing and telemetry-based congestion control, which NVIDIA credits with delivering nearly perfect effective bandwidth plus performance isolation for multi-tenant and multi-job environments. That isolation is what allows a shared cluster to run several jobs without one job's communication pattern degrading another's - the single hardest property to retrofit onto a fabric that was not designed for it. Self-healing network technology handles link failures without a fabric-wide reconvergence.

InfiniBand or Ethernet, at this scale

If you are specifying a pod this large, the fabric choice has usually already narrowed, but the reasoning is worth stating.

InfiniBand's advantages here are collective performance through SHARP, deterministic latency from centrally computed routing, credit-based lossless flow control that does not depend on tuning priority flow control correctly, and topology freedom. The optional router capability extends a cluster across multiple sites, which at ten-thousand-GPU scale is a real constraint rather than a theoretical one.

Ethernet's advantages are operational: the skills, tooling and monitoring already exist, the fabric converges with storage and management, and NVIDIA's own Spectrum-X platform - the SN5610 and SN5600D - now builds seriously for AI on Ethernet with RoCE extensions, adaptive routing and performance isolation of its own.

The honest position is that both paths are now credible and the decision is usually made by existing practice and by what the storage tier demands, not by a benchmark. What does not change is that the out-of-band management network is Ethernet either way - that is the SN2201. The Dell AI fabric switch guide works through the full comparison. Everything at this tier is listed together on our 800G and 400G switch range.

The platform, and the two sibling models

The Quantum-X800 platform is the Q3400-RA, the Q3401-RD and the Q3200-RA switches, the NVIDIA ConnectX-8 SuperNIC and the LinkX interconnect portfolio, achieving end-to-end 800Gb/s from switch to host. Fabric-scale management and monitoring is UFM, which gives software-defined visibility into the performance and health of the network.

The Q3401-RD is the same 4U air-cooled chassis with the same 144 ports, supporting rack cabinets that distribute DC power - the difference is the power feed at 48-54V DC, not the cooling. The Q3200-RA is the smaller-scale option: a 2U enclosure housing two independent switches of 36 ports each at 800Gb/s, which NVIDIA positions for smaller platforms and for connecting new compute clusters to existing Quantum and Quantum-2 storage infrastructure. Neither is listed here today; ask if you need one.

Physical planning, and what NVIDIA does not publish

The chassis is 4U, air-cooled, designed for standard 19-inch rack cabinets, with a 200-240V AC power feed. The control plane is an Intel Coffee Lake 4-core i3-8100H at 3GHz, and security is Integrated Root of Trust across the CPU, CPLD and switch IC. Compliance is CE, FCC, VCCI, ICES and RCM for emissions, and RoHS, CB, cTUVus, CE and CU for product safety.

What the datasheet does not give is as important as what it does. There is no weight, no dimension set beyond the 4U height, no power draw, no thermal output, no latency figure and no operating temperature range published for this model. That is unusually sparse, and nothing here is estimated to fill the gap. Freight quoting, circuit sizing and thermal planning all need either a physical unit or a figure from Dell - ask us and we will get it rather than guess it.

Where it sits in the range

The Quantum-2 QM9700 is the previous generation at 64 ports of 400Gb/s in 1U, internally managed; the QM9790 is its externally managed twin, and the closer operational analogue to this switch since the Q3400-RA is UFM-managed by construction. On the Ethernet side, the SN5610 is the 800GbE Spectrum-X spine, the SN5400 and SN4700 the 400GbE tiers, and the Z9864F-ON Dell's own 800GbE PowerSwitch.

Every unit is professionally reconditioned, fully tested and shipped with warranty. This model is quote-only, and candidly: a flagship announced in 2024 does not exist on the secondary market in any quantity. What we can do is source against a specification. Tell us the node count, topology, splitter plan, cabling split and quantity and a specialist will come back to you on what is achievable and on what timeline.

Specifications