Skip to content

// installation

Reference configuration: 3 × NVIDIA RTX PRO 6000

Applies to

Product: Server · Operation: by the customer or by basebox · Audience: Platform Operator · Hardware: 3 × NVIDIA RTX PRO 6000 · Architecture: Single node

What a reference configuration is

A reference configuration documents a concrete, known setup. It does not mean basebox requires this hardware – the GPUs serve the service models and inference. Shared installation steps are not repeated here; they are in the bare-metal installation guide.

A three-GPU workstation-class configuration, offered by FAST LTA with 128 GB or 512 GB RAM. Typical split: one GPU dedicated to the service models, the remaining two GPUs for inference – individually for two models or as a tensor-parallel pair for a larger model. This is the most affordable of the FAST LTA configurations and designed for mid-sized organisations.

Status

Supported – see Status terms. basebox supports this setup; a measured reference workload as for the Validated configurations does not yet exist.

Architecture

Single Kubernetes node with all three GPUs.

GPU

Physical GPU Role Memory (NVIDIA specification)
RTX PRO 6000 GPU 0 Inference 96 GB GDDR7
RTX PRO 6000 GPU 1 Inference 96 GB GDDR7
RTX PRO 6000 GPU 2 Service models 96 GB GDDR7

Two inference GPUs with 96 GB each are in the VRAM class of the 4 × L40S (192 GB total): documented there are e.g. GPT-OSS 120B AWQ (TP across all inference GPUs) or Llama 3.3 70B FP8 with 32–65k context. Take that as orientation, not measurement – interconnect and memory bandwidth differ.

RAM

128 GB or 512 GB (FAST LTA options).

basebox version

Not yet measured by basebox on this configuration. Use the current release (chart 0.3.32 → basebox 1.8.8).

Tested on

No measurement documented.

Component placement

Component Placement
basebox platform Node CPU
Inference (vLLM) GPU 0 (+ GPU 1 with TP=2 or as a second instance)
GPU RAG, document extraction, OCR, speech-to-text GPU 2

Service model placement

All service models on GPU 2, separate from inference – the Dedicated service GPU pattern. The division of the 96 GB between ragsrv, ragsrv-support and OCR/STT endpoints is pending DevOps confirmation.

Inference configuration

Backend vLLM. Two options: (a) TP=2 across GPU 0 and 1 for a large model with long context; (b) two instances on one GPU each – e.g. a main model and a second one for reasoning or an app – see Multiple inference instances. Guidelines: LLM recommendations.

Tested models

No documented model verification for this configuration.

Deviations from the standard installation

  • Inference requests nvidia.com/gpu: 2 with NUM_GPUS: "2" (TP=2) or two deployments with nvidia.com/gpu: 1 each.
  • Service models request GPU resources on GPU 2; assign GPUs by UUID/PCI address.

Known limitations

  • Workstation class: no NVLink/NVSwitch as with H100 SXM; tensor parallelism over PCIe is possible but slower in the exchange between shards. Check the peer path with nvidia-smi topo -m.
  • Approve the power and thermal budget for three GPUs with the manufacturer or FAST LTA.
  • No high availability across several nodes.