Skip to content

Home / Computing / GPU clusters

GPU clusters

More servers.
One connected system.

Take work beyond a single server. Dedicated GPU servers, high-speed networking and storage are designed and tested as one system in one location.

PLENTYFIELD / CONNECTED
Three graphite equipment cabinets in a fixed perspective studio composition
COMPUTEConnected nodes
LOCATIONOne site
ACCEPTANCEOne complete system

A 32-GPU EXAMPLE

Four nodes.
A dedicated fabric.

Four 8-GPU B300 servers connected within one location. The fabric, switch layout and storage are selected for the workload.

Example topology. The final design and complete server specifications are agreed in the proposal.

32 GPUs / 4 NODES / 1 LOCATIONONE CONNECTED SYSTEM
Four eight-GPU servers. One dedicated fabric.Four dimensional server enclosures represent nodes one through four. Each has eight B300 GPUs. Separate logical routes connect all four nodes to the GPU fabric above and the access and workload storage network below. These lines describe network roles, not physical cabling.DEDICATED GPU FABRICSwitching & topologyACCESS & WORKLOAD STORAGEFront-end networkNODE 018 × B300NODE 028 × B300NODE 038 × B300NODE 048 × B300
DEDICATED GPU FABRIC
NODE 018 × B300
NODE 028 × B300
NODE 038 × B300
NODE 048 × B300
ACCESS & WORKLOAD STORAGE

Compute nodes

Exact GPU configuration, host resources and number of servers.

GPU fabric

Port speeds, switching, topology and any oversubscription, validated for the workload.

Access & storage

Local NVMe, shared storage or a parallel file system, with a route for loading data and checkpoints.

32 GPUs in one connected deployment

Choose the architecture

A cluster does more
than add servers.

For jobs that exchange data between machines, the connections and storage become part of the computing system.

Compare independent servers and a connected cluster
Independent serversConnected GPU cluster
Best suited toInference replicas, separate rendering jobs and work that fits one server.Training and other distributed workloads that exchange data across nodes.
NetworkFront-end Ethernet for access and data transfer.Dedicated inter-node fabric, plus front-end connectivity.
AcceptanceEach server meets the agreed specification.The whole system passes the agreed compute, fabric and storage tests.

Your cluster specification

Every connection counts.

A cluster brings the nodes, fabric and storage together. We assess HGX, AMD Instinct and suitably networked RTX PRO configurations for the work you need to run.

  1. 01
    Compute nodes

    Exact GPU configuration, host resources and number of servers.

  2. 02
    GPU fabric

    Port speeds, switching, topology and any oversubscription, validated for the workload.

  3. 03
    Storage and ingestion

    Local NVMe, shared storage or a parallel file system, with a route for loading data and checkpoints.

  4. 04
    Platform and access

    Operating system, GPU software, access and any separately agreed scheduler setup.

  5. 05
    Acceptance

    Compute, fabric and storage checks, plus your benchmark where agreed.

The reference RTX PRO server with 2 × 25 GbE is a standalone configuration. An RTX PRO cluster requires a separately validated server and network design. Tightly coupled clusters operate within one location.

Cluster pricing

Quoted for the configuration, operating location, service scope and term you choose. Your offer identifies the delivery window and any upfront payment.

What the service includes

Plan your cluster.

Bring the model and the scale. We’ll work through the network, storage and timing with you.

Request a proposal