Skip to content

// installation

Hardware options

Applies to

Product: Server · Audience: Platform Operator · Procurement

Where server hardware comes from: FAST LTA / basebox-supplied systems or hardware procured by the customer. FAST LTA is a hardware partner, not a product of its own and not a deployment model of its own – a FAST LTA system is a basebox Server like any other.

Two sourcing paths

Source through FAST LTA Procure yourself
What you get A pre-configured system that basebox lists as a reference configuration Your own or existing hardware
Status Supported or Validated per configuration Validated if it matches a listed configuration; otherwise Custom – review with basebox
RAM options 128 GB or 512 GB per system As you need
Advantage Known setup, documented GPU allocation, short coordination Use of existing investments, free choice
Effort Low Check against requirements; individual coordination for Custom

Both paths lead to the same software and the same installation procedure.

Current FAST LTA configurations

Configuration GPUs Typical split Status
3 × NVIDIA RTX PRO 6000 3 workstation class Dedicated service GPU, the rest for inference Supported
2 × NVIDIA H200 141 GB 2 data center class 1 inference, 1 as 4 × MIG for services; measured reference workload Validated
2 × NVIDIA H200 + 1 × RTX PRO 6000 2 + 1 2 inference, RTX PRO 6000 dedicated to services Supported

Each with 128 GB or 512 GB RAM; RAM is a property of the configuration, not a configuration of its own. Meaning of the statuses: Status terms.

Self-procured hardware

If it matches a reference configuration, that configuration's status applies – for example 4 × NVIDIA H100 SXM 80 GB (Validated; applies explicitly to SXM, not PCIe or NVL).

If it matches none, it is Custom: possibly functional, but supported only after individual review with basebox. Document the agreed setup with the template Custom hardware.

At minimum, the hardware must meet what the Server Preparation Guide and the chart require:

Component Minimum Recommendation
GPU NVIDIA, compute capability 7.0+ (V100, T4, RTX 20xx, A100, L4, H100 …); driver compatible with CUDA 13.0 (12.9 as fallback) Data center GPUs with enough VRAM for model and service models; MIG-capable if services are to run via MIG
CPU 8 cores 16+ cores
RAM 16 GB 32 GB+, 64 GB+ for large models
Storage 500 GB 1 TB+ SSD/NVMe for models, documents, databases
Network Access to the image registry during installation or offline transfer Separate networks for user access and inference when hosts are separate
Power and cooling For the GPUs at their configured power limit Approved by the server manufacturer

What really decides the GPU question

Not basebox – but inference and service models:

  • Inference: model, quantisation, context length, concurrency and latency target determine VRAM. Orientation: LLM recommendations – e.g. GPT-OSS 120B MXFP4 on 4 × H100 or GPT-OSS 20B on a 24 GB GPU.
  • Service models: RAG, OCR, STT benefit from dedicated GPU capacity (typically useful around 48 GB, better about 96 GB) – as a GPU of their own or MIG slices; functional in CPU mode but slower.
  • Platform: CPU, RAM, storage – no GPU.

A sentence like "basebox needs two H200s" describes an inference decision. Background: Understand the architecture.

Separate hosts

The hardware does not have to be in one machine: a CPU-only application server for the platform (and service models in CPU mode) plus a separate GPU host for inference is a supported layout – see Deployment topologies. That is the path when a GPU system already exists or inference is to sit in a different security zone.

Procurement checklist

  • Target model(s) and expected concurrency known → VRAM need of inference
  • Service models on GPU (which?) or CPU mode decided
  • Configuration checked against the reference configurations; for Custom, review with basebox initiated
  • RAM (128/512 GB or your own value), storage, network sized
  • Power, cooling, rack space approved
  • Location (Hosting options) and operation (Operating models) decided

Next step: Hosting options