NVIDIA Partner · UK
InfiniBand and Spectrum-X Ethernet switches, ConnectX SuperNICs, BlueField DPUs and LinkX cables, designed into a fabric that keeps your GPUs busy.
In a GPU cluster the network is part of the computer. Training jobs spread across many GPUs spend a large share of their time exchanging data, so a slow or poorly designed fabric leaves expensive accelerators waiting. NVIDIA's networking portfolio, which grew out of its Mellanox acquisition, covers both InfiniBand and Ethernet for this east–west traffic.
We design the fabric with the cluster, not after it. That means choosing between InfiniBand and RoCE Ethernet, settling the topology and oversubscription, planning every cable and optic, and proving performance before handover.
Ethernet built for AI clusters
InfiniBand for AI and HPC
Data-centre Ethernet switching
Low-latency switching
Server network adapters
Offload networking, storage and security
The physical layer
Operate, monitor and simulate
Specifications are summarised from manufacturer information and can change between revisions. We don't publish prices here — for current pricing and stock, see the shop.
Build, integration and support from our UK engineering team
We choose between InfiniBand and RoCE Ethernet, design fat-tree or rail-optimised topologies and set oversubscription to match your workloads.
We produce the bill of materials for every cable and transceiver, plan lengths from your rack layout and supply labelled, pre-cut cable kits.
We configure Cumulus Linux, Onyx or UFM, and deploy NetQ or UFM for monitoring.
After installation we run performance tests such as NCCL and ib_write_bw, and hand over the results before go-live.
Trade members get free technical support on every order and unlimited engineer time — help with configuration, drivers, software and troubleshooting, for as long as you own the kit.
Both work well for AI. InfiniBand is a mature choice for dedicated training clusters; Spectrum-X Ethernet suits organisations that want to stay on Ethernet tooling and skills. We compare them against your scale, team and existing network.
A cluster can need hundreds of links, and the wrong length, type or transceiver delays go-live. We plan every link from the rack layout and deliver labelled kits.
Yes. We run standard bandwidth and collective-communication tests and give you the results as part of handover.
Usually, yes. We review what you have and recommend how to extend or replace it.
Tell us what you need to run and where it has to live. An engineer will come back with a recommendation, not a sales script.