// the find
rbonghi/jetson_stats
📊 Simple package for monitoring and control your NVIDIA Jetson [Orin, Xavier, Nano, TX] series
jetson-stats (jtop) is a monitoring and control tool for NVIDIA Jetson boards — think htop but Jetson-aware, exposing GPU, CPU, memory, fan, NVP power models, and JetPack version info. It runs as a system service so you don't need root to read stats from user space or inside Docker containers. The target is anyone deploying ML workloads on Jetson hardware who needs visibility into what's actually happening.
The Unix socket architecture (jtop.sock) is the right call — one privileged service, many unprivileged readers, no sudo in your containers. Hardware detection covers the full Jetson family including Thor, with platform-specific code for CUDA memory and GPU stats where the Tegra stack differs. The Python library API is clean enough to drop into a monitoring script in five lines. Active maintenance with recent commits and structured issue templates for missing hardware/JetPack combinations means the hardware coverage actually grows.
Installation still requires sudo pip3 globally, and the Ubuntu 24.04 path needs --break-system-packages, which is a bad look for anything serious; a proper apt package or venv-compatible install path would fix this. The TUI is built on curses with no headless/JSON output mode for CI or log aggregation — you have to use the Python API separately, so the terminal UI and the library feel like two independent projects. Test coverage appears to require real hardware to run meaningfully (conftest.py likely mocks tegrastats), so it's hard to validate changes without a physical board. No native Prometheus/OpenMetrics exporter — the Flask example in the repo is a starting point, not a solution.