Control Center

Every SkillFishOS tool lives in one window. Open it from the menu (SkillFishOS → Control Center) or with skillfish-control-center. A controller drives it too: D-pad or stick to move, A confirms, B goes back, LB and RB switch section.

The sections

• Status: one number per card: GPU clock and ceiling, CPU, compute units, fan, kernel, governor, Vulkan driver, whether the last boot followed a clean shutdown. Nothing to set.
• Tuner: the voltage/frequency curve of the GPU governor, the CPU, the cores, the compute units, the VRAM split.
• Fan: the fan curve with its lead, the source sensors, the safety limits, the PWM test.
• Monitor: every reading in one chart per unit, plus the per-thread clock bars. REC records, CSV exports.
• Games: our Mesa per launcher or system-wide, the scx_bpfland scheduler, FSR 4, GE-Proton install and default.
• Profiles: ceiling, CPU, fan and scheduler in one click; save your own.
• Kernel: the installed kernels, the default one, boot once, uninstall.
• Snapshots: system snapshots and the btrfs maintenance schedule.
• AI: Unsloth Studio on Vulkan: engine, update, models from the Hugging Face Hub, a chat test, memory and network.
• Emulators: EmuDeck, or the emulators one by one from Flathub.
• Console: Steam Big Picture inside gamescope, now or from the login screen.
• ISO: disk images mounted through udisks.

Tuner

The curve is the chart: MHz across, millivolts up. Drag a knot, double-click to add one, right-click to remove it, or type the numbers in the table beside the chart (+ and − add and remove knots). The dashed vertical line is the ceiling, the blue dot is the GPU right now. Three presets: Cautious 1850 (the sweet spot with the stock heatsink, nearly the same frames and ten degrees less), Balanced 2000, Performance 2100 (the fifteen-point curve measured on the development board, the one we ship).

Apply is a trial. A curve that asks too little voltage hangs the board, and on the BC-250 a hang means pulling the plug. So Apply starts the candidate with the previous curve still on disk and counts down 25 seconds: press Keep and the candidate is written for good; do nothing, or let the board die and come back, and the previous curve is what boots.

The panels open on demand:

• CPU: clock, undervolt step (6.25 mV each) and thermal limit, each with a slider and a number box at step 1. Suggest UV walks the undervolt down two steps at a time with twelve seconds of load per step; Find my max climbs from 3600 to 4000 MHz; Test 60 s loads the current values. Every run shows a countdown and a Stop button; the values under test are never written to disk, so a hang boots the previous ones. Above 3500 MHz with eight cores the gain is nil: heat is in charge.
• Cores: which cores are online, SMT, the 8-core unlock (6c/12t becomes 8c/16t, +20% measured) and Before boot (EFI): the unlock done before GRUB in a single boot, instead of the extra reboot of the in-system service.
• CU: the 40 compute units as 20 cells (pairs): green on, red off, grey kept on by the driver. Type the number of CUs you want (24 to 40, in pairs) or click the cells; Test CU turns the extra pairs on one at a time under vkpeak and reports errors, for the silicon lottery.
• VRAM: the UMA split in the CMOS, applied at the next boot. With the 512 MB dynamic split some games pick low textures: if they look blurry, try 4 or 6 GB fixed. Anything from 512 MB to 12 GB, with presets at 512 MB, 1, 2, 4, 6 and 8 GB.
• Advanced: the governor’s own knobs: ascent margin, step, descent confirmations, thermal and power thresholds, droop.
• Test: vkpeak and the helper log.

Monitor

One chart per quantity, each with its own true scale: temperatures (CPU, GPU, VRM, system, NVMe), clocks (CPU average, min, max, GPU and its ceiling), load, power, voltages, fan and memory (RAM, VRAM, GTT), plus one bar per thread. Click a legend entry to hide a line; hover to read every chart at the same instant. REC writes a .sfmon file (CSV) that Open reloads with a scrubber; CSV exports it for a spreadsheet.

Games

• Vulkan driver: our Mesa opens the compute queues the stock driver keeps closed on this GPU: +4% in Cyberpunk 2077, +12% with FSR 4 on. Per launcher (Steam, Heroic) through a flatpak override, or for the whole system. It needs the SkillFishOS kernel: on another kernel those queues wedge the GPU, and the system-wide switch refuses to start there. The card shows the kernel that is running.
• Scheduler: scx_bpfland, loaded only while a game runs (GameMode raises it at launch and drops it at exit): +1.5-2% in Cyberpunk and steadier frames. If the kernel ejects it twice the service stops until you reset the counter: that is the “Ejected by the kernel” row.
• FSR 4: through OptiScaler on the game’s DLSS path. GE-Proton 11 downloads OptiScaler by itself when it finds PROTON_USE_OPTISCALER=1, the game must be set to DLSS, and AMD’s FSR 4 library (amdxcffx64.dll, from AMD’s Windows driver) goes next to the game. XeSS, where a game has it built in, costs the same as FSR 3 at 1080p.
• Proton: the GE-Proton 11 releases; Install downloads one (about 500 MB) into the Steam and Heroic folders, Default picks it. Steam must be closed for its default to be written.

AI

The engine is Unsloth Studio: it runs GGUF models through llama.cpp on the Vulkan backend, which on the BC-250’s gfx1013 is the only accelerated path, because ROCm does not support it. Measured on the board with Qwen3-1.7B Q4_K_M: 210 tokens per second against 41 on the CPU alone.

• Engine: on, off, at boot. On, it holds GPU memory, so turn it off before gaming. Beside it sit the version, the update check and Update, which reruns the official installer with the Vulkan bundle of llama.cpp.
• Access: Unsloth generates a random password when it installs, and shuts itself down after an hour if that password is not changed. The first sign-in happens here: choose the Studio password and the API key is created with it, and shared with the Remote Manager.
• Models: the ones on disk with quantization and size. To download one you need the repository name on Hugging Face (for example unsloth/Qwen3-4B-GGUF) and the variant picked from the list, sizes included.
• Chat: a quick run over the OpenAI-compatible API (http://127.0.0.1:8888/v1), with the measured tokens per second. The full chat, with files and search, is inside Studio.
• Memory: VRAM, GTT, RAM, swap and the model budget. The GTT limit is a kernel parameter: it takes effect at the next boot.
• Network: open on the local network, Studio answers the other devices at home too, with its own user and password. Parallel chats are llama-server’s slots.

Fan, profiles and the rest

The Fan page draws the curve inside the chart of the sensors: at this temperature the fan turns like this, with a lead so the speed rises before the temperature does. Profiles move the ceiling within the curve that is already there and set CPU, fan preset and scheduler together. Snapshots are pictures of the system, your files are not touched. AI holds GPU memory while on: turn it off before gaming. Console starts Steam Big Picture inside gamescope on top of the desktop, or as a session from the login screen. The Remote Manager mirrors the Tuner, Games and Profiles pages in a browser.