Supported GPUs and accelerators
Stealthium reads GPU telemetry directly from the host and turns it into metrics, events, and security signal. This page covers which GPUs and accelerators Stealthium supports. For operating systems and installation methods, see Minimum requirements and supported versions.
Supported hardware
Stealthium supports:
- NVIDIA data-center and professional GPUs.
- Tenstorrent Wormhole accelerators.
We track the current and previous supported driver branches for each. Older drivers may still work, but they aren't tested and aren't covered by support. We don't pin a minimum driver or CUDA version here — those move with each vendor's release cadence. For the versions validated for your deployment, or to confirm a specific model, contact the Stealthium team.
How Stealthium reads GPU telemetry
On each host, Stealthium collects GPU telemetry from the GPU driver and CUDA stack, including NVML for some data. The agent runs where the driver runs: for GPU passthrough, where the driver runs inside the VM, run the agent in that VM and it instruments the driver there.
DCGM as an additional source
If an NVIDIA DCGM endpoint (dcgm-exporter) is available, Stealthium can scrape it as an additional source, and those samples feed the same metric series. DCGM is optional. See the Metrics Reference for which GPU metrics are exported and where each one comes from.