NVIDIA Q3450-LD (Quantum-X Photonics)
144 × 800G in 4U. That is 115.2 Tb/s of switching through 144 OSFP cages.
- Logical ports144 × 800G
- Physical cages144 × OSFP
- Switching capacity115.2 Tb/s
- Rack units4U
- Typical power3,950 W
- Maximum power3,950 W
- Fabricinfiniband
- ASICNVIDIA Quantum-X (4 chips)
Breakout
A port faster than the endpoint it serves can be broken out when the speeds divide cleanly. That raises the usable port count. Two-tier capacity goes as the square of the port count, so it moves the tier boundary a long way.
| Endpoint speed | Breakout | Usable ports |
|---|---|---|
| 800G | none | 144 |
| 400G | 2:1 | 288 |
| 200G | 4:1 | 576 |
| 100G | 8:1 | 1,152 |
Roles and optics
- Suitable for: compute, storage fabrics
- Cage type: OSFP. A module of any other form factor will not seat
- Accepted optic coding: nvidia
- Management: Internally managed
- Part number:
920-9B36M-00MX-8S0
Coding matters more than it looks. A switch may refuse a transceiver that is not coded for it, and that is one of six axes that have to agree before a link comes up. The others are form factor, reach, fibre type, fibre connector and ferrule polish. See optics compatibility.
Notes
Co-packaged optics, and the first such switch in this catalog. There are no transceiver cages: 144 XDR ports at 800 Gb/s attach fibre directly by MPO, which is why a BOM using it orders no optics for the switch end. That is the point of the design — electrical loss drops from roughly 22 dB to 4 dB because the signal never crosses a pluggable interface, and the switch draws about half what an equivalent pluggable design would. 4RU, liquid-cooled, four Quantum-X chips. UNCONFIRMED: power, and the faceplate fibre connector. MPO-12 and MPO-16 trunks are not interchangeable, so confirm against the NVIDIA cabling guide before ordering.
Source: NVIDIA Q32xx and Q34xx XDR 800Gb/s InfiniBand Switch Systems user manual (docs.nvidia.com/networking/display/XDRSwitchesHWUM) for ports, form factor, chip count, SKU and liquid cooling, verified 2026-09. POWER IS NOT FROM NVIDIA: the ~3.95 kW figure comes from third-party coverage and is unconfirmed. Needs verification