Skip to content

FAQ

Do I need a GPU to develop or test?

No. Tests use a mock collector that returns deterministic data for 2 A100 GPUs. Run make test on any platform.

How does this compare to dcgm-exporter?

dcgm-exporter feeds Prometheus dashboards for humans. gpu-mcp-server gives agents direct tool access — real-time, structured JSON, no query language needed.

Can an agent control or modify the GPU?

No. Every tool is read-only by design. The server observes GPUs and never changes clocks, kills processes, or allocates resources.

What about security and authentication?

stdio transport = local only. The server runs as a child process of the agent with no network endpoint to authenticate. If/when the HTTP transport lands, auth becomes an explicit design item.

Does it support multi-GPU?

Yes. NVML enumerates all local GPUs. list_gpus returns every device.

Does it support multi-node?

Not yet. Today it's per-node (stdio, local process). Multi-node aggregation is on the Q1 2027 roadmap, likely paired with the HTTP transport.

Does it support MIG?

Yes. MIG instances appear as separate devices with is_mig, parent_gpu, and mig_profile fields. An agent can address them individually.

What are the dependencies?

Only two external dependencies:

See DEPENDENCIES.md for details.

What about AMD or Intel GPUs?

AMD ROCm and Intel oneAPI are on the Q2 2027 roadmap. The Collector interface is the seam that makes adding vendors clean — a new backend, same tools and JSON.