VPSie home
Services

GPU servers

GPU nodes that arrive installed and validated, with Kubernetes or Slurm ready. Orchestration covers NVIDIA H200 and GB300 systems.

GPUTrainingInferenceRenderingGPUTrainingInferenceRendering

GPU servers for AI, HPC and rendering

VPSie builds GPU capacity on its bare-metal and orchestration layer: nodes are discovered, installed and prepared for GPU work by automation, then handed to you as a ready cluster. VPSie supports NVIDIA H200 and GB300 systems.

  • Nodes installed over the network with iPXE and Kickstart or Preseed
  • Drivers, container runtime and monitoring installed by playbooks
  • Run Kubernetes, Slurm or a standalone node
  • Controlled through an API as well as the console

GPU node

GPU node

GPU node

GPU node

One pool of GPU nodes

Workloads across the pool

Training
Inference
Rendering

Sample view. Illustrative, not live data.

What the automation does for each node

From a racked machine to a GPU-ready resource.

Provision

Bare-metal provisioning installs the operating system through network boot, using the server management controller (Redfish, IPMI or a vendor interface).

Configure

Playbooks install drivers, the container runtime, shared storage for checkpoints and a monitoring stack, in order.

Validate and register

Validation checks run, results show in the console and API, and passing nodes join the scheduler as GPU-ready.

Two GPU platforms

H200

NVIDIA H200

Hopper generation, 141 GB of HBM3e memory per GPU, up to 4.8 TB/s of memory bandwidth, in SXM and NVL forms.

  • NVLink and NVSwitch bring-up for 8-GPU nodes
  • InfiniBand NDR 400G port checks and subnet manager set-up
  • XID error detection with automatic node quarantine and re-provisioning
  • MIG profiles from 1g.20gb up to 7g.141gb
GB300

NVIDIA GB300

Blackwell rack-scale system with 72 GPUs on NVLink 5.0, running on ARM64 Grace CPUs.

  • Rack-wide Redfish discovery and inventory
  • NVLink fabric validation after boot
  • Liquid cooling health alerts
  • Dynamic MIG changes without a reboot

Software stack ready on day one

Images are prepared for AI work, so the first job does not start with a driver install.

Operating systems

Ubuntu Server LTS (20.04, 22.04, 24.04), Rocky Linux, AlmaLinux, Red Hat, Debian, and NVIDIA DGX OS where the platform needs it.

AI libraries

CUDA, cuDNN and NCCL, with PyTorch, TensorFlow and JAX as supported frameworks.

Containers and scheduling

Docker, Kubernetes with the NVIDIA GPU Operator, or Slurm with GPU-aware resource definitions.

Sharing a GPU

Pass a GPU through 1:1, combine several, or split one with MIG or vGPU so small jobs do not hold a whole card.

Monitoring

DCGM metrics with Prometheus and Grafana dashboards, and automatic alert routing.

Hardware vendors

Works across Dell, ASUS, HPE and Supermicro systems that expose a BMC or API control interface.

FrameworksPyTorch, TensorFlow, JAXContainers and schedulingDocker, Kubernetes + GPU Operator, SlurmAI librariesCUDA, cuDNN, NCCLOperating systemUbuntu LTS, Rocky, Alma, RHEL, Debian, DGX OSGPU server hardwareDell, ASUS, HPE, SupermicroMonitoringDCGM metricsPrometheus andGrafana boardsAutomatic alertroutingShare a GPU: 1:1 passthrough, combine several, or split with MIG or vGPUFrameworksPyTorch, TensorFlow, JAXMonitoringDCGM metricsPrometheus andGrafana boardsAutomatic alertroutingContainers and schedulingDocker, Kubernetes + GPU Operator, SlurmAI librariesCUDA, cuDNN, NCCLOperating systemUbuntu LTS, Rocky, Alma, RHEL, Debian,DGX OSGPU server hardwareDell, ASUS, HPE, SupermicroShare a GPU: 1:1 passthrough, combineseveral, or split with MIG or vGPU
What is ready on the image, from hardware up to frameworks.
GPUGPUGPUGPUGPUGPUGPUGPU12 to 20 minto a working cluster8-node clusterGPUGPUGPUGPUGPUGPUGPUGPU12 to 20 minto a working cluster8-node cluster

Time to a working cluster

Typical deployment time for a full 8-node H200 or GB300 cluster, from bare metal to ready to train, is about 12 to 20 minutes through a single orchestration API call. You pick the GPU type, the stack, whether to turn MIG on, and which nodes take part.

  • One call chooses Kubernetes, Slurm or standalone
  • Validation suite runs before nodes are marked ready
  • Progress and pass or fail results visible in the console
  • Same VPSie account, projects and API tokens

Where it fits

Plans on bare metal

GPU plans appear on bare metal with NVIDIA-ready images. In the beta they are marked beta.

Run it on your site

Use the same orchestration on your own GPU hardware in a private cloud, including air-gapped sites.

Orchestrate with Kubernetes

Add GPU nodes to a Kubernetes cluster with the device plugin and GPU Operator.

Questions about GPUs

Which GPUs are supported?

VPSie supports NVIDIA H200 and GB300 systems. Other models can be listed as plans on bare metal; ask us about the one you need.

Can I run Slurm?

Yes. Slurm and Kubernetes are both supported, as is a standalone node with no scheduler.

Can one GPU serve several users?

Yes. MIG or vGPU partitioning splits a card, and Kubernetes can schedule against the MIG resource labels.

Does it work without Redfish or IPMI?

Yes. Fallbacks include vendor interfaces such as iDRAC and iLO, network boot, agent-based control and hypervisor APIs.

Can I use it offline?

The orchestration does not require an outside service, so it fits air-gapped installs. Software mirrors must be provided on site.

Plan a GPU deployment

Tell us the model, node count and scheduler and we will size it with you.