800 Gb/s per GPU today · 1.6 Tb/s next
Quantum InfiniBand or Spectrum-X Ethernet, SuperNICs, DPUs and cabling: the fabric that decides whether your GPUs work as one machine.
When an AI job runs on more than one server, every step waits for the slowest message between GPUs. A fabric that is merely fast on paper but congested in practice can leave expensive GPUs idle for a large share of every training step. Getting the network right is the cheapest way to get more out of the GPUs you buy.
NVIDIA offers two fabrics for AI factories. Quantum InfiniBand, now at 800 Gb/s with Quantum-X800, is the long-standing choice for the largest training clusters. Spectrum-X brings AI-specific congestion control and adaptive routing to Ethernet, with NVIDIA quoting 1.6 times the AI network performance of off-the-shelf Ethernet, and fits organisations that standardise on Ethernet operations.
We design the whole network estate: the GPU fabric, a separate storage network, the front-end network to your users and the management network, with cabling and optics planned to the metre.
Each GPU's NIC connects to its own rail of leaf switches, so collective operations take the shortest path and traffic stays balanced as the cluster grows.
We weigh your scale, workloads, multi-tenancy needs and your team's skills, rather than defaulting to one camp.
Checkpointing and data loading run on a separate network, often through BlueField DPUs, so they never slow GPU-to-GPU traffic.
NetQ for Ethernet and UFM for InfiniBand give telemetry and fault finding, and feed our SysOps247 monitoring.
Building blocks we design with.
| Ethernet fabric | Spectrum-X switches, including Spectrum-6 based models at 102.4 Tb/s per switch chip (SN6600: 128 × 800G; SN6800: 512 × 800G in 5U) |
|---|---|
| InfiniBand fabric | Quantum-X800 InfiniBand at 800 Gb/s per port |
| SuperNICs | ConnectX-8 (800 Gb/s, PCIe Gen6) for Blackwell; ConnectX-9 (1,600 Gb/s per GPU) for Vera Rubin |
| DPUs | BlueField-3 for storage, security offload and multi-tenant isolation |
| Photonics | Spectrum-X Ethernet Photonics with co-packaged optics: NVIDIA quotes 5 times better power efficiency than pluggable optics |
| Cabling | LinkX DAC, AOC and transceivers, with a labelled cabling plan |
| Operations | NVIDIA NetQ (Ethernet), UFM (InfiniBand), integrated with your monitoring |
Figures are NVIDIA's published specifications for the platform. Where NVIDIA quotes two Tensor Core figures, the first assumes sparsity and the second is dense. The exact build (CPUs, memory, storage, cooling, networking) is agreed with you and confirmed on a written quote.
Network items that are easy to forget.
| Space and power | Leaf, spine and storage switches need rack space and power of their own |
|---|---|
| Cable routes | Overhead or underfloor pathways sized for hundreds of high-speed cables |
| Optics budget | Transceivers can be a significant share of fabric cost; we size them precisely |
| Growth | Spine ports and pathways reserved for the next racks |
One accountable UK team from sizing to day-two operations
Rail-optimised fabric, storage, front-end and management networks designed together.
Switches, NICs, optics and cables counted to the port and the metre.
Cabling, firmware alignment and NCCL all-reduce benchmarks before handover.
Telemetry, congestion tuning and fault response as part of operations.
NVIDIA data-centre GPUs are allocated by NVIDIA and the server makers, and lead times are measured in months rather than weeks. Complete, tested systems usually get allocation sooner than loose parts, so we quote these platforms as built systems. Shipments outside the UK, and many advanced GPUs within it, need an end-user and end-use check under UK and US export rules before we can confirm supply.
Trade members get our engineers' time on sourcing and design questions, early stock and waitlist priority, and free technical support on every order.
Both work well for AI. InfiniBand has the longest track record for very large training. Spectrum-X suits multi-tenant factories and teams that run Ethernet everywhere else. We recommend based on your scale and skills.
Usually for the front-end and management networks, rarely for the GPU fabric. Blackwell systems expect 400 to 800 Gb/s per GPU, which older switches cannot supply.
Tell us what you want it to produce, how many people will use it and where it would live. An engineer will come back with a sized recommendation and an honest view on whether owning makes sense for you.