THE OMBRAHOST COLLECTION / GPU compute

GPU computeMake room for bigger ideas.

Single-accelerator NVIDIA configurations with 24, 48 or 80 GB of GPU memory and separate host-memory allocations. Match inference, rendering or selected training workloads to the actual memory budget and supported image.

3configurations
4compatible destinations
$929.10monthly plans from
Explore the configurations
THE DETAILS COME FIRST

Find your fit.

See the accelerator, video memory and supporting resources together.

3 configurationsComplete inclusions. Full cycle totals.
GPU compute

Spectre 4090

Pair one NVIDIA RTX 4090 with 24 GB of GPU memory, 64 GB of host memory and 1 TB of NVMe storage for inference, rendering and experiments. CUDA image options and a 10 Gbps uplink support your software stack; model fit depends on precision, context length and runtime overhead.

  • AcceleratorOne NVIDIA RTX 4090 with 24 GB VRAM
  • Compute8 vCPU (Threadripper)
  • Memory64 GB DDR5
  • Storage1 TB NVMe SSD
  • Transfer20 TB / month
  • DDoS protection40 Gbps
  • IPv41 included
  • Network port10 Gbps
  • PlatformBare-Metal
  • IPv6/64 subnet
  • One NVIDIA RTX 4090 with 24 GB VRAM
  • Locations limited to Iceland, Switzerland, the Netherlands and Moldova
  • PyTorch, vLLM and Ollama stack compatibility
  • 10 Gbps uplink
Locations & software

Ubuntu 22.04 + CUDA 12 · Debian + CUDA 12 · Custom

  • GPU memory is separate from the 64 GB host-memory allocation
  • CUDA image options are listed in the operating-system selection
  • A 13B FP16 model needs about 26 GB for weights alone and exceeds 24 GB VRAM without a different memory strategy
  • Quantisation, context length, KV cache and runtime overhead determine model fit
$929.10/ month

Billed monthly

Configure plan View full specifications
GPU compute

Spectre A6000

Use one NVIDIA RTX A6000 with 48 GB of GPU memory alongside 128 GB of ECC host memory and 2 TB of NVMe storage. The larger GPU memory allocation and 10 Gbps uplink suit more demanding inference or selected training workflows, with model and optimiser memory assessed before deployment.

  • AcceleratorOne NVIDIA RTX A6000 with 48 GB VRAM
  • Compute16 vCPU (EPYC)
  • Memory128 GB DDR5
  • Storage2 TB NVMe SSD
  • Transfer30 TB / month
  • DDoS protection80 Gbps
  • IPv41 included
  • Network port10 Gbps
  • PlatformBare-Metal
  • IPv6/64 subnet
  • One NVIDIA RTX A6000 with 48 GB VRAM
  • ECC host memory
  • Selected training and inference workflows with model-dependent sizing
  • 10 Gbps uplink
Locations & software

Ubuntu 22.04 + CUDA 12 · Custom

  • GPU memory is separate from the 128 GB host-memory allocation
  • A 13B FP16 model needs about 26 GB for weights alone; gradients, optimiser state and runtime memory are additional
  • CUDA image options are listed in the operating-system selection
  • Training feasibility depends on precision, batch size, optimiser and memory strategy
$2,088.10/ month

Billed monthly

Configure plan View full specifications
GPU compute

Spectre H100

Build intensive accelerator workloads around one NVIDIA H100 with 80 GB of HBM3, 256 GB of host memory and 4 TB of NVMe storage. A 25 Gbps uplink and 50 TB monthly transfer allowance support large datasets; high-parameter models still require careful precision, context and memory planning.

  • AcceleratorOne NVIDIA H100 with 80 GB HBM3 GPU memory
  • Compute32 vCPU (EPYC)
  • Memory256 GB DDR5
  • Storage4 TB NVMe SSD
  • Transfer50 TB / month
  • DDoS protection100 Gbps
  • IPv41 included
  • Network port25 Gbps
  • PlatformBare-Metal
  • IPv6/64 subnet
  • One NVIDIA H100 with 80 GB HBM3 GPU memory
  • Large-model workflows subject to memory sizing
  • 25 Gbps uplink
  • 99.99% catalogue availability target
Locations & software

Ubuntu 22.04 + CUDA 12 · Custom

  • GPU memory is separate from the 256 GB host-memory allocation
  • A 70B FP16 model needs about 140 GB for weights alone; one 80 GB H100 requires quantisation, offloading or a suitable multi-device strategy
  • NVLink concerns a separately arranged multi-GPU configuration; this plan specifies one accelerator
  • CUDA image options are listed in the operating-system selection
$5,508.10/ month

Billed monthly

Configure plan View full specifications

A configuration. A destination. Your choice.

Explore deployment regions and practical guides alongside the exact requirements of your project.

Explore use cases
A LITTLE MORE CLARITY

Before your
next move.

Practical details for GPU compute, from setup to the choices ahead.

Open the knowledge centre

These are prepared service specifications. Infrastructure, licences and optional services require validation before a live launch.

GPU COMPUTE / IN DETAIL7 useful answers
Which OmbraHost GPU configurations are prepared?+

The Spectre range lists an RTX 4090 with 24 GB VRAM, RTX A6000 with 48 GB and H100 with 80 GB. Model weights, context cache, batch size and software overhead all affect fit. Locations and images vary by tier. Installed models, measured token rates and GPU allocation require a provisioning integration.

How much GPU memory is needed?+

FP16 weights need roughly two bytes per parameter; four-bit weights roughly half a byte. Allow additional room for context cache, batching and runtime overhead.

What are the GPU choices?+

Spectre 4090 specifies 24 GB VRAM, A6000 48 GB and H100 80 GB. Benchmark the actual model and precision before choosing.

Will a 70B FP16 model fit?+

Its weights alone need roughly 140 GB. A single 80 GB H100 cannot hold them; evaluate quantisation or a suitable multi-GPU configuration.

Can I request multiple GPUs?+

Custom multi-GPU arrangements are a prepared sales enquiry. Stock, NVLink topology, power and lead time must be confirmed before any commitment.

Are short research bursts available?+

The source describes a seven-day minimum by arrangement. This remains a proposed commercial option; the current checkout uses the three published billing cycles.

Can Windows be used?+

Only the catalogue templates appear in the standard selector. Windows GPU or custom image requests require driver, licence and deployment eligibility checks.

A little more flexibility.

PREPARED OPTIONS

Explore optional services and their billing units. Optional service requests require a connected provider before fulfilment.

Veil Edge

Prepared network profile with the 10 Gbps mitigation allocation, anomaly detection and latency tuning. Validate its relationship to protection already included in the plan.

$4.75per month

Eclipse Shield

Prepared 100 Gbps profile with application rules and a dedicated filtering path. Confirm effective capacity and compatibility; some base plans already specify greater capacity.

$19.00per month

Iron Veil

An engineer-led baseline review covering CIS guidance, login protection, key handling and firewall rules. This service is prepared and has no assigned engineer yet.

$46.55one-time

Vault Mirror

Prepared daily incremental off-site copies, seven-day retention and an encrypted S3-compatible destination. The destination and recovery procedure require production validation.

$4.75per month

Checkpoint

Prepared scheduled and manual snapshots with a rolling five-snapshot retention limit. This fixed service fee is separate from the proposed per-GB storage fee.

$2.85per month

First Light

A proposed initial setup covering installation, security baseline and agreed basic services. The one-time scope must be confirmed before work begins.

$19.00one-time

Shadow Transfer

A prepared transfer service included at no charge with an eligible new order. Inventory, access, cutover and rollback are agreed before migration.

$0.00one-time

Route Extension

One additional IPv4 address, subject to allocation eligibility and technical approval. Selecting the option does not reserve or route an address.

$1.90per month