A CPU and GPU monitor with gadget-style instruments. Each card has a dual dial (load plus an inset temperature sub-dial, after the Vista CPU meter gadget), a Geiger-counter activity strip with a flashing LED, and the detail beneath it.
Glass (default), Matrix, Hive, Windows, macOS, Linux, Dark and White. Pick one from
the toolbar, or start with one using ?skin=matrix. Every skin is a block of CSS
tokens at the top of web/index.html; the dials and strips read the same tokens,
so adding a skin means adding one block and one <option>.
Download PCmeter-<version>-linux-x86_64.tar.gz from the
latest release, then:
tar xzf PCmeter-*-linux-x86_64.tar.gz
cd PCmeter-*-linux-x86_64
./install.shThat puts PCmeter in ~/.local/bin and adds an app-menu entry and a desktop
shortcut; nothing outside your home folder changes. On KDE Plasma, the first click on
the desktop icon may ask whether to run it. ./install.sh --uninstall removes it all.
The shortcut runs pcmeter --app, which shows the CPU and GPU cards as desktop gadgets: each card is
its own frameless, translucent window with the desktop blurred behind it, and there is no outer
window. Drag a card anywhere to move it, drag its left or right edge to change its width, and hover
it for its menu and dismiss buttons. The menu (also in the system tray) picks the skin, the view
(Full, Compact, or Gauges only), clicks, keep-on-top, which cards are shown, and Launch on startup
(starts PCmeter when you log in). In the browser window (--app --browser, and on Windows and macOS)
right-click anywhere for the same checkbox. It is per user and needs no admin rights: an entry in
~/.config/autostart on Linux, the Run key under HKEY_CURRENT_USER on Windows, and a LaunchAgent
in ~/Library/LaunchAgents on macOS. Launching PCmeter again brings back any card you dismissed. On KDE Plasma (Wayland) the cards return to where you left
them. NVIDIA cards need the proprietary driver; AMD cards work out of the box.
The gadgets are native Qt Quick (qml/, no browser engine). They draw with Qt's hardware renderer (OpenGL ES), at up to 24
frames a second with film-style motion blur and only while something moves. By default each card is half the height of its screen's work area (the screen minus panels), so two stacked
cards fill it, with the content scaled to fit; a width you set by dragging an edge is kept as saved, and the
default follows a card to a screen of another size. A fresh start stacks the cards in one column at the right edge,
processor on top.
Rendering: PCMETER_RENDER=cpu switches to the software renderer for hosts with no usable GL. The default
(gpu) uses OpenGL with Qt's basic render loop and __GL_YIELD=USLEEP (set unless you have set it), because the
NVIDIA driver busy-waits in eglSwapBuffers: on an RTX 5080 under KWin Wayland the two cards cost about 11% of one core
this way, against 28% with the threaded loop. PCMETER_RENDER=threaded restores the threaded loop and
PCMETER_RENDER=vulkan uses the Vulkan backend; QSG_RENDER_LOOP and QSG_RHI_BACKEND still win if set.
While the session is locked, the cards are hidden or all dismissed, the sampler sleeps on an event and
nothing is drawn. The process lists are scanned every 5 s, and only while a card's list is open. pcmeter --app --browser
opens the old single-page app window in Chrome, Chromium, Brave, Edge or Vivaldi instead.
Clicks: each card's activity strip has a speaker button that mutes or unmutes that card (cards start muted) and a volume slider on hover; the toolbar's Clicks button (in the gadgets, the menu) is the master switch.
pip install -r requirements.txt
python3 backend/pcmeter.py --openThe UI is served at http://127.0.0.1:8765. Add ?demo to see simulated data, or
open web/index.html directly from disk.
Compact hides the detail and puts a 60-second load graph in each card's header. Gauges shows only each card's name and dial.
| Flag | What it does |
|---|---|
--interval 0.5 |
Seconds between samples |
--amd-peak-bw 960 |
Peak VRAM bandwidth in GB/s for AMD cards, so throughput shows in GB/s |
--temp-source auto |
Dial temperature: auto/coolant use the radiator/coolant sensor when there is one, cpu always the die |
--liquidctl PATH |
liquidctl executable for AIO coolers without a kernel driver (default: found on PATH; none disables). Only liquidctl status is ever run |
--fanctl-state PATH |
fanctl state JSON behind the fan field (default: /run/fanctl/state.json, then /run/host/run/fanctl/state.json; none disables). Ignored when older than 15 s or missing |
--port 8765 / --host |
Where to listen |
--app / --app --browser |
Desktop gadgets, or a browser app window |
-
History:
GET /historyreturns{"interval": 5, "samples": [[t, coolant, die, gpu0], ...], "events": [{"t", "on", "trigger"}, ...]}, 5-second averages of the last 2 hours (epoch seconds, deg C,nullwhen missing; oldest first) and fan-override transitions. It is kept in memory and only fills while a window or stream is open. Every snapshot also carriesfanfrom fanctl's state file, ornullwithout fresh state:{"override", "trigger", "duty", "rpm", "radiator": {"duty", "rpm"}, "pump": {"duty", "rpm", "stalled"}}. Fans are read by name (radiator,pump);duty/rpmstay the radiator's, and an older state with a single fan of any name still works (pumpis thennull). The CPU card showsRadiator 1,710 rpm · 52 %andPump 1,820 rpm; a stalled pump readsPUMP STALLin red, and the Temps chart labels an overridefans 100% · pump stallwhentriggerispump. -
CPU: total and per-thread load, clock, temperature (
k10temp,coretemp, ...), load average, package power (RAPL; often needs root to read). A coolant/radiator temperature (hwmon: Kraken, D5 Next, Aquacomputer, ...; elseliquidctl status) is reported ascoolant_tempand preferred for the dial when present. An AIO with no coolant sensor (e.g. Arctic Liquid Freezer III Pro) just leaves itnull: the dial shows the die, no coolant labels appear, and an auto-detectedliquidctlthat finds no liquid temperature stops polling after a few tries. -
The dials. The CPU dial shows plain CPU utilisation (busy %, averaged over 3 s in the desktop app). The backend still works out how close the chip is to whichever limit it meets first, the largest of up to three terms, each 0-100 %: POWER (package watts / PL1 from
intel-raplconstraint_*_power_limit_uw; watts from RAPL, else fanctl'sinputs.power), THERMAL ((die - 40 C) / (TjMax - 40 C), TjMax from hwmontemp*_crit, else 100 C) and LOAD (utilisation); a missing source drops its term. THROTTLED lights on the CPU dial while the package throttle counter (thermal_throttle/package_throttle_total_time_ms) rose within the last 3 s. The NVIDIA dial shows that closest limit itself (power / enforced limit, temperature vs the slowdown threshold, NVML's throttle reasons), labelled with the winning term. The snapshot carries it aslimit(value,term,terms,throttled,power_limit_w,temp_limit). The CPU "Package power" row reads75.4 / 120 W. -
NVIDIA (NVML): load, temperature, clocks, power, fan, VRAM, top 10 processes with per-process load and VRAM.
-
AMD (amdgpu sysfs): the same, with per-process data from
/proc/<pid>/fdinfo. -
VRAM throughput is an estimate: memory-controller busy % x bandwidth at the current memory clock (clock x bus width x 2), shown in MB/s or GB/s. The busy % is averaged over every driver sample between updates. NVML has no direct bytes-per-second counter, and its GPM DRAM counter reads zero on GeForce cards without profiling rights. The graph scales to recent traffic.
-
Activity is a Poisson process whose rate rises steeply with load, from about 25 counts per minute at idle to about 4,800 at full load. Clicks plays them as audio, CPU on the left channel and GPU on the right.
pip install -r requirements.txt pyinstaller
packaging/build.sh- Desktop widget: wrap
web/in Tauri with a transparent, frameless, always-on-top window and load it with?widget(transparent background). - Windows: psutil and NVML already work; CPU temperature needs LibreHardwareMonitor.
- macOS: psutil works; temperature and GPU need IOKit /
powermetrics.