Licensed to be used in conjunction with basebox, only.
// installation
Reference configurations
Applies to
Product: Server · Audience: Platform Operator
A reference configuration documents a concrete, known hardware setup: its topology, GPU allocation, configuration differences and the software/model combinations it has been exercised with.
It does not mean "this is what basebox requires". basebox itself needs little GPU; the GPUs in these configurations serve service models and inference. Hardware changes quickly – which is why these pages live inside the Server documentation rather than forming its structure, and why new GPUs or multi-node layouts can be added here without touching the navigation.
Current configurations
| Configuration | Status | Architecture | System RAM | Notes |
|---|---|---|---|---|
| 3 × NVIDIA RTX PRO 6000 | Supported | Single node | 128 GB / 512 GB | FAST LTA offering |
| 2 × NVIDIA H200 141 GB | Validated | Single node | 128 GB / 512 GB | Measured reference workload; MIG service GPU |
| 2 × NVIDIA H200 + 1 × RTX PRO 6000 | Supported | Single node | 128 GB / 512 GB | FAST LTA offering; dedicated service GPU |
| 4 × NVIDIA H100 SXM 80 GB | Validated | Single node | see page | Tensor-parallel inference pair + MIG services |
| Custom hardware | Custom | — | — | Requires individual verification |
128 GB vs. 512 GB system RAM is a property of a configuration, not a separate configuration. FAST LTA offers both for each of its systems.
Status terminology
| Status | Meaning |
|---|---|
| Validated | basebox has tested this exact configuration. |
| Supported | basebox supports this setup, even if every possible combination has not been benchmarked recently. |
| Experimental | Expected to work, but not production-validated. |
| Custom | Customer-specific configuration that requires individual verification. |
Do not claim that arbitrary hardware is validated. If a setup is not listed, it is Custom until basebox has verified it.
What a configuration page contains – and what it does not
Every page follows the same template:
Status · Architecture · GPU · System RAM · basebox version · Tested on · Component placement · Service model placement · Inference configuration · Tested models · Required deviations from standard installation · Known limitations · Related pages
A configuration page documents only what is specific to that hardware: values, topology, GPU allocation, Helm/config differences, limitations, tested combinations.
It does not repeat the installation procedure. Common steps – OS, drivers, Kubernetes, storage, networking, installing basebox, validation – live once, in the bare-metal installation guide, and every configuration page links there.
Conceptual allocation pattern
Most configurations follow one pattern; exact GPU-to-service mapping is stated per page only where DevOps has confirmed it.
flowchart LR
subgraph INF["Inference GPU(s)"]
LLM["LLM inference<br/>complete GPU(s), optionally tensor-parallel"]
end
subgraph SVC["Service GPU (dedicated GPU or MIG slices)"]
RAG["GPU RAG / embeddings"]
EXT["Document extraction"]
OCR["OCR"]
STT["Speech-to-text"]
end
P["basebox platform<br/>(CPU, no GPU)"]
P --> SVC
P --> INF
style INF fill:#dbeafe,stroke:#1e40af,color:#1d1e1c
style SVC fill:#dcefe2,stroke:#3a7a49,color:#1d1e1c
style P fill:#f4f2ee,stroke:#524e47,color:#1d1e1cAdding a configuration
Copy Custom hardware as the template, fill every section, assign a status honestly, and add one row to the table above. No navigation change is needed.
Next step: Bare-metal installation