gpur

btop-style GPU monitor for the terminal. An nvtop replacement that runs everywhere: NVIDIA, AMD, Intel, and Apple GPUs on Linux, macOS, and Windows.

Mirrored braille waveforms, gradient meters, an nvtop-style process table with per-process GPU attribution, mouse support, foldable GPU cards, and theming via the hjkl stack.

Status: beta. The AMD Linux path is battle-tested on real hardware, and the macOS IOKit backend returns live data (validated against macOS CI runners' paravirt GPU — Apple Silicon numbers still welcome). Intel and Windows backends are implemented, CI-exercised, and awaiting real-GPU reports.

Install

cargo install gpur              # crates.io
yay -S gpur-bin                 # Arch (AUR)
brew install kryptic-sh/tap/gpur  # macOS
scoop install kryptic/gpur      # Windows

Debian/RPM packages, musl builds, and an Alpine .apk are attached to each GitHub release.

Run

gpur              # real GPUs
gpur --mock       # fake GPUs, works anywhere
gpur --mock 6     # demo a 6-GPU rig
gpur --once       # one text snapshot, no TUI (scripts, quick checks)
gpur --json       # machine-readable snapshot (waybar/polybar, monitoring)
gpur --log x.jsonl  # append one JSON record per poll (benchmark/sensor logs)
gpur --graphs ascii # graph glyphs: braille (default) | block | ascii

Keys & mouse

Configuration

gpur reads $XDG_CONFIG_HOME/gpur/config.toml (falls back to ~/.config/gpur/config.toml). No file is written automatically; missing file means built-in defaults.

tick_ms = 1000
history_len = 300
# graph glyphs: "braille" (default), "block", or "ascii" for fonts
# without braille coverage
graphs = "braille"
# hjkl-theme TOML; omit for the built-in theme
theme = "/path/to/theme.toml"

Themes use the hjkl-theme schema: [palette] of hex colors plus [ui] styles. Any hjkl theme file works — the meters, waveform gradients, and highlights all derive from the palette.

Backends

Probe order: nvml → amdgpu → intel → ioaccel → pdh. Intel utilization is derived from per-client fdinfo engine counters (i915 busy-ns, xe cycles) the same way nvtop does it; power comes from the hwmon energy counter delta. PDH is the vendor-generic Windows fallback (Task Manager's counters) and reports utilization and memory only.

Backend poll failures degrade gracefully: the last snapshot stays on screen with a header warning until polling recovers — a driver reset won't kill the monitor.

Extras: each GPU card shows session peaks/averages (util, temp, power) when tall enough, and the PCIe caption flags a link running below its maximum (PCIe 3.0@8x (max 4.0@16x)) — a classic symptom of a bad riser or wrong slot.

Encoder/decoder activity shows in each card (split enc/dec on NVIDIA/Windows, unified video on AMD/Intel), and a red throttle badge appears when the GPU is thermal- or power-limited (real reasons on NVIDIA, at-limit heuristic on AMD).

Planned depth: Apple temperature/power (SMC/IOReport), Windows AMD ADLX (temps/fans/clocks).

Per-process GPU attribution

The process table (PID, user, device, type, GPU%, GPU memory, CPU%, host memory, command) works on Linux (AMD + Intel via /proc fdinfo — same-user processes unless run as root), NVIDIA everywhere (NVML), and Windows (PDH per-pid counters). macOS has no public per-process GPU API. On Linux, containerized processes get a CONTAINER column (docker/podman/k8s id from cgroups).

Recorded sessions replay: gpur --log run.jsonl while it happens, then gpur --replay run.jsonl anywhere — including machines with no GPU — with the recorded process attribution intact. Ideal for bug reports.

License

MIT