IT EN
Usage

Native apps — Tuner and AI

SkillFishOS's graphical tools for controlling hardware and AI without a terminal.

SkillFishOS includes two native applications (written in PyQt6, themed with Kvantum) that put hardware and AI-stack control in the user’s hands without touching the terminal.

SkillFishOS Tuner

The Tuner is the hardware control panel. It lets you adjust:

  • CPU overclock and undervolt;
  • GPU safe-points (via the SMU governor, see GPU and overclock);
  • the fan (PWM control);
  • the UMA VRAM (requires a reboot);
  • the Compute Units, live — see below.

Live Compute Units (grid)

The Tuner shows the GPU’s CUs as a grid of squares (4 SE/SH rows × 5 WGP): green = active, red = off. You toggle them live, no reboot — click the pairs (1 WGP = 2 CU) or use the 24 / 32 / 40 CU presets — then Apply. The first 24 CUs are the driver minimum and stay always on (see GPU and overclock).

SkillFishOS Tuner — the live Compute Unit grid, presets and CU test

CU test (silicon lottery)

The “CU test” button checks the health of the extra CUs: it enables each pair alone, stresses it with vkpeak and watches for GPU faults/hangs, plus a final full-40 stress. It’s there to catch defective CUs on salvaged/“discard” APUs, so you know whether your chip sustains all 40 CUs.

CU test result — all pairs OK, 40 CUs stable at 11380 GFLOPS, no defects

“Test” flow and live monitor

The “Test” flow (CPU, GPU, CU, fan): apply a change → run a benchmark → verify stability and, if something is wrong, perform an automatic rollback. When any test starts, a Monitor window opens with live charts of temperature, frequency, voltage and fan (closable at will).

Tuner Monitor window — live temperature, frequency, GPU voltage and fan charts during a test

Architecture: a user GUI plus a small root daemon that performs the privileged operations. On a personal PC it is configured not to ask for a password on every operation. The desktop HUD also shows the active CUs live.

SkillFishOS AI

The AI panel turns the local LLM stack on and off with one click, freeing GPU and RAM for games when not needed. It’s the “easy” front-end of the stack described in On-device AI.

SkillFishOS AI panel — local LLM engine (Qwen3 14B) on the Vulkan GPU, on/off with one click

Why they exist

SkillFishOS’s goal is that anyone — including the youngest — can use and tune the system without having to learn terminal commands. These apps translate complex operations (SMU governor, kernel parameters, Docker containers) into a few clicks, while keeping the safeguards (thermal-guard, test-and-rollback) always active.

Sources