all-smi

all-smi monitors GPUs, NPUs, CPUs, and memory from one terminal — on your workstation or across a cluster. It also exports Prometheus metrics and machine-readable snapshots for monitoring systems and scripts.

• Local and cluster views: live utilization, memory, temperature, and power, with interactive sorting and filtering.
• Multi-vendor monitoring: NVIDIA, AMD, Intel, Apple Silicon, and accelerator-specific backends.
• Remote access: HTTP exporters or SSH, plus cluster-wide user and topology views.
• Automation: Prometheus, JSON/CSV snapshots, recording/replay, and a Rust library API.

Local-node view (on macOS)

All-node view (remote mode)

Node view (remote mode)

Supported hardware

Not every device exposes every metric. See hardware requirements for drivers, permissions, and limitations.

Installation

Homebrew (macOS / Linux)

brew install all-smi

Ubuntu PPA

sudo add-apt-repository ppa:lablup/backend-ai
sudo apt update
sudo apt install all-smi

Other installation methods

• Pre-built binaries and Debian packages are available from GitHub Releases.
• Cargo: cargo install all-smi — see build prerequisites. Cargo does not install the Linux AMD companion library; use a release archive, package, or source build for Linux AMD monitoring.
• Build from source.

On glibc Linux, keep liball_smi_amd.so with the downloaded binary for AMD monitoring. The main binary can run without it. No container image is published.

Quick start

Monitor this machine

all-smi

With default settings this opens the local TUI. all-smi local selects it explicitly. Press h for help or q to quit. macOS does not require sudo; other platforms may need device permissions for particular readers.

Monitor remote nodes

Start an exporter on each node:

all-smi api --port 9090

Connect from your workstation:

all-smi view --hosts http://node1:9090 http://node2:9090

Metrics are available at http://node1:9090/metrics. Keep exporters on a trusted network; use your firewall or an authenticated proxy to control access. For machines without an exporter, see SSH monitoring.

Export or capture metrics

# One-shot JSON (with default configuration)
all-smi snapshot

# Capture 30 seconds, then replay in the TUI
all-smi record --output trace.ndjson.zst --duration 30s
all-smi view --replay trace.ndjson.zst

See snapshots and recording for CSV, scripting, configuration caveats, and capture limitations.

Check your environment

all-smi doctor
all-smi config path
all-smi --help

Configuration is optional. To keep an exporter running across reboots, follow the service guide for your platform and installation method.

Recent Updates

• v0.28.0 (2026/10/17): Added NVIDIA RTX Spark N1X and native Windows ARM64 support with DLA metrics on the GPU row, read the N1X DLA engine clock, and fixed false 100% Intel Xe utilization from frozen GT idle counters.

Documentation

These documents track main. For a released version, browse the documentation at its Git tag and consult the release history.

Contributing

Bug reports, hardware compatibility reports, and pull requests are welcome. Start with the contribution guidelines. For missing devices or metrics, see how to collect a diagnostic report.

Acknowledgments

The project's development background and tooling are described in the development story.

License

Apache License 2.0.