Networking infrastructure for modern compute environments.

Compute is easy to quote in isolation and hard to deploy that way. Port speeds, fabric choice, optics and cable plant decide whether the systems work as a cluster.

Not sure what to specify?

Start with what needs to connect, the number of systems, expected speed, and any equipment that must remain in place.

Top-of-rack switching for a server refresh
A fabric for a GPU cluster
Adapters and optics for an existing environment

What we can source

  • Top-of-rack and spine switching
  • High-speed Ethernet, from 10/25G through 100/200/400G
  • InfiniBand fabrics for tightly coupled clusters
  • Server NICs, SmartNICs and DPUs
  • Optics, transceivers, DACs and cabling
  • Management and out-of-band networking

Ethernet or InfiniBand

For most general compute and storage networks, Ethernet is the practical default: broad supply, familiar operations and a wide range of speeds. For tightly coupled training clusters and some HPC workloads, InfiniBand is chosen for latency and collective communication behaviour.

The decision is rarely purely technical. Operational familiarity, existing switching, support model and supply availability all count. Where it is genuinely open, stating that invites both options to be quoted.

Details that make a network quote accurate

ItemWhat to state
Port count and speedPer rack and in total, including uplinks.
OversubscriptionAcceptable ratio between access and uplink capacity.
Optics and cablingDistances involved, and whether optics are in scope.
NIC placementOCP slot or PCIe, and how many slots are available.
RedundancyDual-homing, redundant switching, failure domain expectations.
GrowthNode count today and the count the fabric must reach.

How 46 Systems handles networking requirements

Networking is sourced alongside the compute it serves, so port counts, NICs and optics are reconciled against the systems rather than quoted in isolation. Requirements stay manufacturer-neutral unless an existing environment makes a specific platform necessary.

What helps us quote accurately

  • Quoting switches without the optics and cables needed to use them
  • Overlooking uplink capacity while sizing access ports
  • Assuming NIC slots are free in a fully populated server
  • Sizing the fabric for today's node count only
  • Mixing fabric types across a cluster without an operational reason

Describe your networking requirement.

Switching, NICs and fabric, alongside the compute it supports.