Skip to content

// installation

Advanced architectures

Applies to

Product: Server · Audience: Platform Operator

Layouts beyond a single node with one inference GPU. These pages exist so the information architecture does not change when such setups get documented – and so it is clear which pattern is validated today, which is supported and which is future. The yardstick is the same as for the reference configurations: Validated · Supported · Experimental · Custom.

The four patterns

Pattern What it solves Status today Page
Multi-GPU – tensor parallelism across several GPUs for one model Models or contexts that do not fit on one GPU Validated in the 4 × H100 SXM configuration (TP=2) Multi-GPU
Dedicated service GPU – one GPU or MIG slices exclusively for service models Uploads and transcriptions never compete with inference Validated as MIG variant (2 × H200, 4 × H100); Supported as its own GPU (2 × H200 + RTX PRO 6000) Dedicated service GPU
Multiple inference instances – several models or instances side by side Default plus reasoning model; more concurrency at the same context Possible via endpoints offering several models; routing details pending DevOps confirmation Multiple inference instances
Multi-node – distribute platform, service models and inference across nodes Capacity beyond one server, security zones, high availability Separate application and inference hosts documented and proven (Deployment topologies); high availability is a separate architecture Multi-node

When to read on here

  • The desired model does not fit on one GPU → Multi-GPU.
  • Users report latency spikes in the chat while documents are processed → Dedicated service GPU.
  • Two models are to be offered in parallel, or one instance is saturated → Multiple inference instances.
  • One server is not enough (GPU slots, power), inference belongs in another zone, or resilience is required → Multi-node.

Everything before that – one node, one inference GPU, service models on MIG or CPU – is covered by the reference configurations and the bare-metal path.

What applies to all patterns

  • Assign GPUs by stable identity (UUID, PCI address), never by index numbers.
  • GPU memory is local per card. Two 80 GB GPUs are not one 160 GB device; tensor parallelism distributes layers, it does not pool memory.
  • Users never talk to inference directly – not even with several instances or nodes. Only AISRV.
  • One model cache per node. Model artefacts live where inference runs.
  • No silent fallback. If an instance or node fails, basebox shows the error.
  • Measure first, then build. Every pattern starts with a capacity or latency measurement – Monitoring, Scaling.

Reading status honestly

Listing a pattern as Supported or Experimental means: basebox knows and supports the approach but has not measured every combination. If you run such a layout in production, document it following the Custom hardware template and report your experience to support@basebox.ai – that is how patterns become reference configurations.

Next step: Multi-GPU