NVIDIA Quantum-3 Q3200-RA
NVIDIA Quantum-3 Q3200-RA
“Crocodile” – Air-Cooled Two-Adjoining XDR Switch
A 2U, dual-ASIC InfiniBand switch built for AI and HPC fabrics that need 800Gb/s XDR bandwidth without the footprint of a full 4U chassis. In this post we break down the full data sheet, the numbers that matter, and how to get a quote.
Two switches, one 2U box, zero compromise
The Q3200-RA is NVIDIA’s compact answer to the Quantum-X800 platform: instead of one large switch ASIC filling a 4U chassis, NVIDIA pairs two independent Quantum-3 switch chips side by side in a single 2U enclosure. Each chip runs its own 36-port, 18-cage InfiniBand fabric, so the box behaves electrically like two switches sharing power, cooling, and a rack unit – hence “two-adjoining.” For teams who don’t need the full 144-port density of the Q3400 family but still want native 800Gb/s XDR bandwidth, this is the sweet spot: enough radix to build a meaningful leaf layer or a small standalone cluster fabric, in half the rack space and roughly half the power draw of the larger chassis.
It’s built to sit comfortably next to earlier-generation gear too. Because each port can be configured down to 400, 200, 100, 56, or 40Gb/s, the Q3200-RA is a practical bridge switch for teams migrating GPU clusters onto XDR while their storage fabric is still running Quantum or Quantum-2 InfiniBand. You get a forward-compatible uplink layer today, and a straightforward path to full 800Gb/s end-to-end once the rest of the fabric catches up.
Headline specifications
XDR vs. the previous InfiniBand generations
Every InfiniBand generation has roughly doubled per-port bandwidth. XDR is the current ceiling, and it’s what the Quantum-3 ASIC inside the Q3200-RA speaks natively.
Full technical specification
| Attribute | Specification |
|---|---|
| Model | Q3200-RA (Quantum-X800 platform) |
| Manufacturer Part Number | 920-9B34F-00RX-FS0 |
| Switch ASIC | NVIDIA Quantum-3, two independent chips per chassis |
| Architecture | Two-adjoining – two 36-port switches in one enclosure |
| Total Ports | 72 XDR ports (36 per switch) |
| Physical Cages | 36 OSFP cages (18 per switch), each splittable for breakout |
| Max Port Speed | 800 Gb/s per port (XDR), 64b/66b line encoding |
| Supported Speeds | 40 / 56 / 100 / 200 / 400 / 800 Gb/s via breakout cabling |
| Aggregate Switching Capacity | 57.6 Tb/s (28.8 Tb/s per switch ASIC) |
| Form Factor | 2U rack-mount, standard depth, rail kit included |
| Airflow / Cooling | C2P (connector-to-power, reverse) airflow, air-cooled |
| Power Supplies | 4× hot-swappable PSUs (AC power cords not included) |
| Management | Dedicated in-band InfiniBand management port; NVIDIA Unified Fabric Manager (UFM) support |
| In-Network Computing | 4th-generation NVIDIA SHARP (Scalable Hierarchical Aggregation and Reduction Protocol) |
| Routing | Adaptive routing with telemetry-based congestion control; optional router mode for multi-site cluster expansion |
| Reliability | Self-healing network technology, performance isolation |
| Operating Temperature | Up to 35°C ambient |
| Storage Temperature | Down to -40°C |
| Certifications | CB, cTUVus, CE, FCC, RoHS, EMC |
| Typical Use Case | Compact XDR leaf/spine layer; bridging new GPU clusters to Quantum / Quantum-2 storage fabrics |
What the platform actually gives you
SHARP In-Network Computing
Fourth-generation SHARP offloads collective operations – reductions, all-reduce, broadcasts – onto the switch fabric itself, cutting the data volume that has to traverse the network during distributed training jobs.
Adaptive Routing
Traffic is rebalanced across available paths in real time, so a single congested link doesn’t become a tail-latency problem for the whole job.
Telemetry-Based Congestion Control
The fabric watches queue depth and flow behavior continuously, throttling senders before congestion turns into dropped or retransmitted traffic.
Self-Healing Fabric
Link and node failures are detected and routed around automatically, which matters most during the multi-day runs typical of large-scale training.
Optional Router Mode
The same hardware can be configured to route between InfiniBand subnets, useful when a cluster needs to grow across multiple physical sites or pods.
RDMA Throughout
Remote direct-memory access keeps CPU overhead out of the data path, which is the difference between a fabric that keeps GPUs fed and one that bottlenecks them.
Where the Q3200-RA sits in the Quantum-3 XDR lineup
NVIDIA offers the same Quantum-3 ASIC across four fixed-configuration switches. The Q3200-RA is the smallest and the only one built around two adjoining 36-port switches rather than one larger fabric.
| Model | Chassis | Ports / Cages | Power / Cooling |
|---|---|---|---|
| Q3200-RA | 2U | 72 ports / 36 OSFP | 4× PSU, air-cooled (C2P) |
| Q3400-RA | 4U | 144 ports / 72 OSFP | 8× PSU, air-cooled (C2P) |
| Q3401-RD | 4U | 144 ports / 72 OSFP | 48VDC busbar |
| Q3450-LD | 4U | 144 ports / 144 MPO | 48VDC, liquid-cooled (CPO) |
If your rack budget or power envelope can’t accommodate a 4U switch, or you’re standing up a smaller GPU pod, the Q3200-RA gets you native XDR bandwidth without over-buying radix you won’t use this cycle.
Request a quote for the Q3200-RA
Send us your configuration and quantity, and we’ll come back with pricing, lead time, and cabling recommendations for part number 920-9B34F-00RX-FS0.
Pricing varies by quantity, cabling package, and support tier. Lead times are confirmed at quote. Please replace the placeholder email above ([email protected]) with your team’s sales address before publishing.
