Creative Compute
0

HPE HDR InfiniBand switches and adapters

ProductConfigure to orderManufacturer documented

The dedicated cluster network. A 1U switch with forty ports, each running at 200Gb/s, plus the adapter cards that go in each server. This is what lets GPUs in different machines train one model together.

Ask Carl about this

No payment is taken online. Every request is an enquiry answered by an engineer.

Representative system image

Where does this sit?

  1. 01Local
  2. 02Workstation
  3. 03Departmental
  4. 04Enterprise
  5. 05Cluster
  6. 06Rack-scale

AI Factory & Cluster. Several GPU servers joined by a dedicated fabric with shared storage, operating as one pool of capacity. Typically bought by an enterprise ai platform team or a research institution. Datacentre.

Fit

Is this the right thing for you?

Who it's for

  • Organisations building a multi-server GPU cluster
  • HPC sites extending an existing InfiniBand fabric
  • Research groups running distributed training

What it runs

  • Distributed model training
  • HPC message-passing workloads
  • Low-latency cluster interconnect

When it is the wrong answer

  • A single-server deployment — you do not need a fabric for one machine
Where it goes
Datacentre rack, alongside the compute it connects
Complexity
specialist
Cooling
Air cooled
Specification

What the numbers mean

Every figure below is explained in plain English. Switch to the technical view for the bare specification.

Switch

Form factor

1U switch

Ports

40x QSFP56

Port speed

HDR 200Gb/s per port

MeansEach server link runs at 200 gigabits per second.

Aggregate bandwidth

Up to 16Tb/s non-blocking

MeansEvery port can run at full speed simultaneously.

Matters“Non-blocking” is the specification that stops the fabric becoming the bottleneck.

Port latency

Sub-130ns (underlying Mellanox QM8790 datasheet)

WhenDistributed training, where latency compounds across every synchronisation step.

Adapters

Adapter options

Single-port 200Gb HDR, or dual-port up to 200Gb aggregate

Supported servers

HPE ProLiant XL and DL series

Verification

Manufacturer documented — HPE InfiniBand collateral a00062184enw and adapter QuickSpecs a00062185enw. Reviewed 2026-08-31.

Manufacturer source
Before you commit

Practical considerations

The things that catch people out after the hardware has already been ordered.

  • Port count sets the ceiling on cluster size. Design for where you intend to be, not where you are.
  • Cabling is the commonest source of delay. We will produce the cable schedule as part of the quotation.
Completeness

Nothing arrives working on its own

What else will I need?

  • One adapter per server, matched to the switch generation
  • Cables or optics of the correct length and type — this is where most orders go wrong
  • A topology design if the cluster will grow
AI networkingAI storageInstallationAsk Carl what I’m missing
Goes with

What normally sits alongside it

Relationships documented by the manufacturer, or by us during a deployment.