Compute Units (40 CU)

Compute Units (40-CU)

The BC-250's RDNA 2 GPU physically has 40 Compute Units, but the driver boots with only 24 active. SkillFishOS can run all 40.

What it buys you

Configuration	vkpeak fp32 (GFLOPS)
tkg + governor, 24 CU	6,868
tkg + governor + 40 CU	11,329

≈ 1.8× the compute throughput. More CUs = more graphics/compute power, but also more heat and power draw.

Two ways it's enabled

1. Boot-time (kernel cmdline): amdgpu.bc250_cc_write_mode=3 makes active_cu_number go from 24 → 40 at GPU init. The Tuner flips this for you and reboots. It's a boot-time setting with no runtime toggle by itself.
2. Live, no reboot: SkillFishOS routes up to 40 CUs at runtime through umr, restored at boot by the skillfish-cu.service systemd unit. You can toggle and test CUs while the system is running.

Manage them from the Tuner

The SkillFishOS Tuner has a live Compute Unit grid:

• Each square = 1 CU — green = active, red = off.
• Click a pair to toggle it (CUs are managed 2 at a time = 1 WGP).
• Presets: 24 / 32 / 40 CU, applied live with «Apply live».
• The first 3 pairs per row (24 CU) are the driver minimum and stay always on.

Or from the CLI: skillfish-cu get / skillfish-cu <mask>.

CU health test (silicon lottery)

Not every chip is stable with all 40 CUs. The Tuner's CU health test enables each extra pair (WGP3–4) on its own, stresses it with vkpeak, and checks for GPU errors/hangs — so you can find a bad CU instead of chasing random crashes. It takes ~2–3 minutes and loads the GPU.

See Benchmarks for what the extra CUs do in real workloads.