CPU & GPU Tracker

A project by Qalarc — AI systems studio, Sydney, Australia. Built, shipped and maintained by Qalarc.

GTK4 compute monitor for AMD/ROCm boxes — per-process CPU with runaway detection, amdgpu VRAM/GTT utilization from sysfs, rocm-smi telemetry, and the zombie-process hunting that grew out of a real incident.

Python 3GTK4 / PyGObjectpsutilamdgpu sysfsrocm-smihwmon

CPU & GPU Tracker is the compute-side companion to RAM Tracker, built for a machine where the GPU is a first-class citizen. Per-process CPU accounting comes from psutil, but the interesting parts are the platform-specific layers: temperatures through hwmon with a friendly-name table for the actual chips on the board (k10temp, amdgpu, nvme), and GPU state read directly from amdgpu sysfs — VRAM and GTT usage, engine busy percentage — with rocm-smi as the authoritative source for power and clocks.

Its signature feature was born from a real incident: two runaway ripgrep processes silently accumulated 298 hours of kernel CPU each over 28 hours while the machine looked 'idle'. The tracker now detects kernel-time-dominated processes — high sy% with low visible activity — and flags them before they cook the machine. Zombie reaping status rounds out the picture.

Deliberately Linux-native: the GPU/thermal panel is amdgpu/ROCm-specific and proud of it. Cross-platform coverage belongs to the Tauri monitors; this one goes deep instead of wide.

Capabilities

Per-process CPU usage with kernel-time runaway detection (the sy% signature)
AMD GPU VRAM/GTT utilization and busy percentage from amdgpu sysfs
rocm-smi telemetry: temperature, power, clocks
Sensor friendly-names for real board chips (k10temp, amdgpu, nvme, acpitz)
Zombie process detection and accounting
Sustained-load process highlighting for agent/compile workloads

CPU & GPU Tracker

Capabilities

Tags