06 — GPU Governor

Which Governor Should You Use?

Installation: cyan-skillfish-governor-smu (Recommended)

Fedora

sudo dnf copr enable filippor/bazzite
sudo dnf install cyan-skillfish-governor-smu
sudo systemctl enable --now cyan-skillfish-governor-smu.service

Bazzite (rpm-ostree)

sudo dnf copr enable filippor/bazzite
rpm-ostree install cyan-skillfish-governor-smu
systemctl reboot
# After reboot:
sudo systemctl enable --now cyan-skillfish-governor-smu.service

CachyOS (paru pre-installed)

paru -S cyan-skillfish-governor-smu
sudo systemctl enable --now cyan-skillfish-governor-smu.service

Arch / Manjaro (install an AUR helper first)

# Install paru (if not present)
sudo pacman -S --needed base-devel git
git clone https://aur.archlinux.org/paru.git
cd paru && makepkg -si && cd ..

# Then:
paru -S cyan-skillfish-governor-smu
# Or with yay: yay -S cyan-skillfish-governor-smu
sudo systemctl enable --now cyan-skillfish-governor-smu.service

Debian / Ubuntu

wget https://github.com/Magnap/cyan-skillfish-governor/releases/latest/download/cyan-skillfish-governor-smu_amd64.deb
sudo dpkg -i cyan-skillfish-governor-smu_amd64.deb
sudo systemctl enable --now cyan-skillfish-governor-smu.service

Generic Linux (Build from Source)

git clone https://github.com/filippor/cyan-skillfish-governor.git
cd cyan-skillfish-governor
# See repo README for build instructions (SMU variant is on smu branch)

How the Governor Works (Technical Overview)

Understanding the governor's internal behavior helps tune it for your specific board and cooling setup (explained by filippor, May 2026).

Timing and Sampling

Every adjust microseconds (default: 200,000 µs = 200ms) the governor runs a sampling cycle:

1. Samples the GPU's busy-flag 64 times, once every sample µs (default: 500 µs)

2. Calculates the percentage of time the GPU was busy vs idle

3. Compares to the load_target range (upper/lower thresholds)

4. If GPU utilization is above upper target → increase frequency to next safe-point

5. If GPU utilization is below lower target → decrease frequency to previous safe-point

6. Every flush-every cycles (default: 10) → updates the displayed frequency (visible in MangoHud)

Timeline:
  |-- adjust (200ms) --|-- adjust (200ms) --|
  | 64x sample (500µs) | 64x sample (500µs) |
  | busy-flag check    | busy-flag check    |
                   ↕ (every 10 cycles = flush-every)
                   MangoHud frequency updates

Safe-Point Interpolation

Between two safe-points, the governor uses the voltage of the next higher safe-point (filippor). Examples:

This means the lowest safe-point [1000, 800] controls voltage for ALL frequencies between 0-1000 MHz. A conservative GPU that never exceeds 1800 MHz during gaming only needs safe-points from 1000-2000 MHz range.

Load Target Tuning

The load_target controls how aggressively the governor scales frequency (filippor, May 2026):

Why a lower upper target means MORE aggressive: The governor tries to keep GPU utilization below the upper target. With upper=0.65, it will ramp frequency up sooner (at 65% utilization). With upper=0.80, it waits until 80% before ramping. This is counterintuitive — the lower the upper target, the more the GPU boosts.

Metrics Bug Workaround

On some kernels, GPU metrics reporting is broken. Try setting fix-metrics = false in config.toml if you see incorrect GPU utilization in MangoHud (community report, May 2026).

Decky Loader Plugin: GPU Clock Control

Project: mix3d/bc250-perf-profile-switcher ↗

A Decky Loader ↗ plugin for BC-250 systems. Features:

Install via Decky Loader's plugin store or manually from the GitHub repo. Ideal for living-room setups where you want on-the-fly GPU control from a controller-friendly interface.

Configuration

Config file: /etc/cyan-skillfish-governor-smu/config.toml

Recommended Config (Good for Most Users)

safe-points = [
    [1000, 700],   # 1000 MHz @ 700 mV (idle)
    [1500, 900],   # 1500 MHz @ 900 mV
    [2000, 1000],  # 2000 MHz @ 1000 mV (gaming)
    [2100, 1025],  # 2100 MHz @ 1025 mV
    [2175, 1050],  # 2175 MHz @ 1050 mV (overclock)
    [2300, 1075],  # 2300 MHz @ 1075 mV (OC, good air cooling)
]

[load_target]
min = 0.70    # 70% GPU utilization target minimum
max = 0.95    # 95% GPU utilization target maximum

[timing]
interval_ms = 50       # How often to check GPU load
burst_samples = 20     # Samples before jumping to max frequency

Conservative Config (Lower Temps / Less OC)

safe-points = [
    [1000, 700],
    [1500, 900],
    [2000, 1000],
]
[load_target]
min = 0.70
max = 0.95
[timing]
interval_ms = 50
burst_samples = 20

Idle Power Testing (May 2026)

Community testing shows CU count has minimal impact on idle power:

Downclocking the GPU to 10 MHz works, but 1 MHz crashes the kernel driver (pops1cl, May 2026). The default minimum frequency in the governor config has been bumped from 400 MHz to 500 MHz.

After Changing Config

sudo systemctl restart cyan-skillfish-governor-smu

SMU-Plus Config Options (Emerging)

The cyan-skillfish-governor-smu-plus variant adds these config options:

# /etc/cyan-skillfish-governor-smu-plus/config.toml
fix-metrics = true            # Enables GPU usage metrics fix
set-method = "smu"            # "smu" or "kernel"
[frequency-range]
min = 400                     # Minimum allowed frequency (MHz)
max = 2200                    # Maximum allowed frequency (MHz)
[dbus]
enabled = false               # D-Bus service

Note: SMU-plus has known issues on CachyOS — the regular SMU governor is more stable.

8-Core GPU Clock Reporting Fix (Aug 2026)

Unlocking all 8 CPU cores breaks GPU clock reporting (pp_dpm_sclk shows nonsense like 1: 15Mhz instead of 1500Mhz — the GPU still clocks correctly, it's a reporting bug). Two fixes exist:

1. fix-freq = true (filippor, commit be9537f, Aug 2026) — fixes reporting in userspace via bind mounts, no kernel patch needed. Add it next to fix-metrics in the governor config:

   fix-metrics = true
   fix-freq = true

⚠️ Arch repo version may lag behind — if fix-freq isn't available, build from the filippor/cyan-skillfish-governor repo directly (e.g. yay -S --editmenu cyan-skillfish-governor-smu, update version/URL, stub b2sums with SKIP — hexxeh, Aug 2026).

2. Kernel patch (keroppl_wizard) — see 02-bios-and-firmware.md 8-core metrics fix options.

Reported working on 8-core boards — GPU frequency reads correctly again (community reports, Aug 2026; hexxeh and lordantares confirmed the config option and the AUR build workaround on kernel 7.1-era CachyOS; dizzey0709 confirmed the updated governor restores GPU freq reporting "without kernel changes", Aug 9 2026).

Governor v0.4.0 (May 2026)

Released to the bc250-collective organization. Key new feature: CPU-based control of memory clocks — lowers memory controller and Infinity Fabric clocks at idle for additional power savings (codyrainy, May 2026). See bc250-collective/cyan-skillfish-governor ↗.

Manual Testing

To test a specific frequency/voltage combination before adding it to the config:

# Stop the governor
sudo systemctl stop cyan-skillfish-governor-smu

# Manually set frequency (in MHz) and voltage (in mV)
echo "vc 0 2100 1050" | sudo tee /sys/devices/pci0000:00/0000:00:08.1/0000:01:00.0/pp_od_clk_voltage

# Commit changes
echo "c" | sudo tee /sys/devices/pci0000:00/0000:00:08.1/0000:01:00.0/pp_od_clk_voltage

# Run a game or benchmark (30+ minutes for stability test)

# If stable, add to your config
# If it crashes, increase voltage or lower frequency

# Restart governor when done
sudo systemctl start cyan-skillfish-governor-smu

SMU Performance Profile Tuning (Power Savings)

Changing the SMU performance profile index can reduce idle power:

# Profile 3 (default): ~75W idle
# Profile 2: ~65W idle
# Profile 1: ~60W idle (disputed — most users report 60-70W floor)
# Profile 0: ~60W idle (gennro reports ~60W from 75W, saving ~15W)

Note: perfprofileindex works on BIOS v3 but NOT on BIOS v5. This is an advanced tuning step — test stability carefully.

Known Stable Frequency / Voltage Points

40 CU voltage guidance (May 2026):

(source: governor.md:577-583)

40 CU Unlock Without Kernel Patch

As of May 2026, 40 CU unlock no longer requires a kernel patch. Use bc250-cu-live-manager ↗ by vinnijs.dev — an interactive TUI that toggles CUs on the fly via UMR. No kernel modifications needed. Works on stock kernel across all distributions including Bazzite (auto-detects dri path).

git clone https://github.com/WinnieLV/bc250-cu-live-manager.git
cd bc250-cu-live-manager
# Follow README for your distro

The tool persists across reboots via systemd (bc250-unlock.service). As of June 2026, the live-manager also supports disabling stock WGPs (Work Group Processors) — useful for isolating defective CUs during artifact troubleshooting. See 10 — Troubleshooting for binary search approach to finding bad CUs.

For kernel parameter-based masking: amdgpu.disable_cu=X.Y.Z uses WGP-pair indexing. See 02 — BIOS & Firmware for full 40 CU documentation.

CPU Overclocking (Advanced)

Uses the bc250_smu_oc ↗ tool to raise CPU boost ceiling:

git clone https://github.com/bc250-collective/bc250_smu_oc.git
cd bc250_smu_oc
pip install --user .

# Test frequency (auto-tunes voltage)
sudo bc250-detect -f 3900 -v 1280 -k

# Make permanent
sudo bc250-detect -f 3900 -v 1280 -k -c /etc/bc250-overclock.conf
sudo bc250-apply -a -i /etc/bc250-overclock.conf
sudo systemctl enable bc250-smu-oc

Verified Results (Fedora 43, Kernel 6.19.8)

(source: governor.md:610-619)

Async Compute Queue Fix (GFX1013, Aug 2026)

By DryhoppedIPA. Linux hard-disables the BC-250 GPU's dedicated compute queues (ACE, the hardware behind async compute) — enabling them on stock software corrupts frames. Root cause is three separate problems: (1) the chip powers on with ACE dispatches in threadgroup-dimension mode mis-executing PARTIAL_TG_EN (RADV's internal copy shaders hit this, corrupting their own copies — the exact defect was known since the GCN3 era and never added to GFX1013's workaround list), (2) stock kernels break the compute-queue lifecycle (repeated teardown wedges the queue), and (3) RADV misdetects the chip as RDNA2 instead of RDNA1-class.

The fix is a 3-patch kernel side + 3-patch Mesa/RADV side:

32-bit games fallback fix (keroppl_wizard, Aug 8 2026): Originally cidhems reported "+10% extra on Expanse Telltale with kernel-mesa patch in Debian; not working on 32bits games" — keroppl_wizard fixed the 32-bit fallback after that: "alright i fixed it. 32bit fallback should just work now. also made it build in release mode" and "robocop rogue city picked up some fps now" (keroppl_wizard, Aug 8 2026).

PS5GPU-BC250 (GUI Controller)

A visual Qt-based GPU controller like MSI Afterburner for Windows:

ACPI Fix (CPU Power States)

The bc250-acpi-fix ↗ enables proper CPU idle and frequency scaling:

Confirmed working on kernel 6.19.8. Both tables loaded via initrd override. See the bc250-acpi-fix repo for full installation instructions. (source: governor.md:640-706)

Verification

Check Governor is Running

systemctl status cyan-skillfish-governor-smu
# Should show: active (running)

Check Frequency Scaling

cat /sys/class/drm/card1/device/pp_dpm_sclk

# Example output:
# 0: 1000Mhz
# 1: 1500Mhz
# 2: 2000Mhz *
#
# The * indicates active frequency

Test dynamic scaling: check frequency at idle (~1000 MHz), start a game, check frequency under load (2000+ MHz).

Monitoring Tools

CoolerControl (GUI):

# Fedora
sudo dnf copr enable terra/terra
sudo dnf install coolercontrol

# Bazzite
ujust install-coolercontrol

MangoHud (In-Game Overlay):

sudo dnf install mangohud  # Fedora
sudo pacman -S mangohud    # Arch

# Steam launch option: mangohud %command%

(source: governor.md:294-322)

Performance Comparison

(source: governor.md:504-511)

Dépannage & Résolution des Problèmes

Governor Won't Start

# Check service status
sudo systemctl status cyan-skillfish-governor-smu

# Check logs
sudo journalctl -u cyan-skillfish-governor-smu

# Enable and restart
sudo systemctl enable cyan-skillfish-governor-smu
sudo systemctl restart cyan-skillfish-governor-smu

GPU Stuck at 1500 MHz

Check: governor is running, minimum voltage is not below 700 mV, config file exists.

# Check governor status
systemctl status cyan-skillfish-governor-smu

# Check minimum voltage
cat /etc/cyan-skillfish-governor-smu/config.toml

# Check kernel frequency range
cat /sys/devices/pci0000:00/0000:00:08.1/0000:01:00.0/pp_od_clk_voltage

Black Screen During GPU Reset

If the governor causes crashes during GPU-intensive games, temporarily disable it:

sudo systemctl stop cyan-skillfish-governor-smu
# Play the game, then re-enable
sudo systemctl start cyan-skillfish-governor-smu

Long-term fix: use stable voltage/frequency settings that don't cause GPU crashes. Increase voltage or reduce max frequency for problematic games.

Governor High CPU Usage

Normal CPU usage: 0.9-1.3%. If usage exceeds 2%, increase polling interval:

# /etc/cyan-skillfish-governor-smu/config.toml
[timing]
interval_ms = 100  # Increase from 50

(source: governor.md:482-502)

Community Resources

(source: governor.md:711-716)

Source of truth: elektricM docs/system/governor.md, docs/system/power.md, docs/bios/overclocking.md

Last updated: 2026-06-14