Reference

Performance & benchmarks

Every real BC-250 benchmark on SkillFishOS — screenshots, full settings, clocks, voltages, temperatures and power.

This is the complete benchmark section: every run was done on our own BC-250 with SkillFishOS, with real screenshots, all the settings used, and the telemetry of clocks, voltages, temperatures, power and fan recorded during the run.

⚠️ Silicon lottery + cooling. The numbers hold for this chip with adequate cooling. The stock cooler is marginal: “back-to-back” comparisons without pauses are skewed by heat-soak — let the card cool a few minutes between runs.

Test conditions (test rig)

Apply to all benchmarks below unless stated otherwise.

ItemValue
BoardAMD BC-250 — Zen 2 “Oberon” + RDNA 2 “Cyan Skillfish” APU (gfx1013)
Memory16 GB GDDR6 unified (UMA)
Compute Units40 / 40 active (routed live, see GPU)
Kernel7.0.10-skillfishos (linux-tkg)
DriverMesa 26.0.8 — RADV (Vulkan) / radeonsi (OpenGL), ACO
GPU governorcyan-skillfish — idle 350 MHz / 700 mV, load 2230 MHz / ~1000 mV
OC profileTurbo/Crazy (GPU cap 2230 MHz, CPU 3.9–4.0 GHz)
Thermal cap85 °C (SMU + thermal-guard), fan on auto
Resolution1920×1080

Architecture reminder: CPU and GPU share the same die and the same power budget. Under mixed load the CPU spontaneously gives up clock (≈3.4–3.5 GHz) to stay within budget and under 85 °C — not a defect, the chip self-protecting.


🎮 Black Myth: Wukong — 112 FPS (1080p)

Black Myth: Wukong — 112 FPS average at 1080p on the AMD BC-250

SettingValue
Resolution1920×1080
FPS capnone (uncapped)
Load typeCPU / draw-call bound
UpscalingFSR 4 unavailable (RDNA 4) → gamescope FSR1/NIS or OptiScaler

Result: average 112 FPS · max 128 · min 92 · 1% low 101.

Telemetry during the run (~4 min):

MetricMeasured value
GPU clock~1.4–1.6 GHz (not saturated: the game is CPU-bound)
GPU edge83–86 °C
GPU power~90–140 W
GPU mV~970–987 mV
CPU clock~3.5 GHz (dropped from 3.9 due to the shared budget)
CPU temp85 °C (at the cap)
VRAM~1.9 GB (menu) → ~4.4 GB (in game)
Fan~2950–3140 RPM

Lesson: in gameplay on a draw-call bound title like Wukong, what matters most is CPU stability under load and good cooling.

Governor Balanced vs Performance (benchmark tool)

The benchmark tool’s flythrough is GPU-bound, so there the clock matters. Switching the governor to Performance in the Tuner (it holds the GPU at its top safe-point under load, idling to 350 MHz):

Governor modeAverage5%-low
Balanced (default)100 FPS92 FPS
Performance111 FPS102 FPS

+11% on the average and on the slowest frames, just from holding the clock high. For safety the Tuner caps the GPU at 2200 MHz @ 1000 mV with a multi-point voltage curve: 2230 MHz at 1000 mV is undervolted and can hard-freeze the machine.


🧪 Unigine Superposition — 1080p HIGH: 12938

Unigine Superposition 1080p High — score 12938 on the BC-250

SettingValue
Version1.1
Graphics APIOpenGL
Resolution1920×1080, fullscreen
ShadersHigh
TexturesHigh
DOFenabled
Motion Blurenabled

Result: score 12,938 · FPS min 75.59 · avg 96.77 · max 127.16. Config read by the tool: CPU AMD BC-250 @ 3894 MHz, RAM 7 GB, GPU AMD BC-250 8 GB (Cyan Skillfish), kernel 7.0.10-skillfishos.


🧪 Unigine Superposition — 1080p EXTREME: 5513

Unigine Superposition 1080p Extreme — score 5513 on the BC-250

SettingValue
Version1.1
Graphics APIOpenGL
Resolution1920×1080, fullscreen
ShadersExtreme
TexturesHigh
DOFenabled
Motion Blurenabled

Result: score 5,513 · FPS avg 41.25 (min ~32.8 · max ~49).

Unigine Superposition — scene rendered in real time A Superposition scene rendered in real time on the BC-250.


🏔️ Unigine Heaven 4.0 — 113.7 FPS · score 2865

Unigine Heaven 4.0 — 113.7 FPS, score 2865 on the BC-250

SettingValue
Graphics APIOpenGL
Resolution1920×1080, windowed
Anti-aliasing
QualityUltra
TessellationExtreme

Result: 113.7 FPS · score 2865 · min 54.8 · max 219.5. Platform read by the tool: Linux 7.0.10-skillfishos x86_64 · CPU AMD BC-250 ×12 · GPU gfx1013.

Unigine Heaven — scene rendered in real time The Heaven scene rendered in real time on the BC-250 during the run.


⚙️ GPU compute — vkpeak (synthetic)

Vulkan compute throughput on the same board, before and after unlocking the 40 CUs.

MetricBaseline 24 CUSkillFishOS 40 CU
FP32 scalar6141 GFLOPS11,329 GFLOPS (11,385 cold)
FP16 vec412,26022,685
int8 dot-product24,550 GIOPS45,495 GIOPS
FP64 scalar385~640
copy d2d (internal bandwidth)191 GBPS

With the 40 CUs active: +85% FP32 over baseline (≈11.3 TFLOPS). When hot, under sustained stress, it settles around 10,214 GFLOPS. At idle the governor drops to 350 MHz, edge ~54 °C after the load.

📦 Memory bandwidth — clpeak

MetricValue
Measured GDDR6 bandwidth~350–367 GB/s
mclk adjustableNo (fixed memory clock)
Memory seen by Vulkan~13 GiB (with extended GTT)

🔧 Tuner profiles — clocks, voltages, temperatures

ProfileCPUCPU voltageGPUPeak temp
Stock (ISO default)3500 MHz1500 MHzthe lowest
Performance3700 MHz~1106 mV (scale −16)2000 MHzbalanced
Turbo3900 MHz~1199 mV (scale −24)2230 MHz< 85 °C (cap)
Crazy4.0 GHz~1224 mV (scale −36)2230 MHz~83 °C in 120 s stress
  • Hard maximum Vid: 1.325 V (never exceeded).
  • 85 °C thermal cap on all profiles; fan on auto; at idle the GPU sits at 350 MHz / 700 mV.

🌡️ Thermal validation (stress test)

Data recorded during the Tuner’s automatic validation (test-and-rollback).

PhaseClockTemperatureNotes
IdleCPU ~2.5 GHz · GPU 350 MHzk10 46 °C · GPU 45 °Cat rest
CPU stress (12 threads, 120 s)CPU 3.68–3.69 GHzk10 85 °C (at the cap)GPU stays at 350 MHz/56 °C
GPU stress (vkpeak loop, 120 s)GPU 2000 MHzedge up to 86 °Cat 86 °C the governor drops to 1819–1900 MHz (thermal-guard); the CPU falls to ~2.2–2.4 GHz due to the shared budget

📊 Comparisons

Same hardware, just changing the OS — Superposition 1080p Extreme on the same BC-250:

SystemScore
SkillFishOS (GPU 2230 · CPU 3900, 40 CU)5513
Other distro (Bazzite, stock clocks)4102

+34% real performance from the very same chip, thanks to 40 unlocked CUs, a governor pushing 2230 MHz and CPU OC+undervolt.

Versus desktop Radeons (Superposition 1080p High): the BC-250 with SkillFishOS (12,938) matches an RX 6600 / 6600 XT costing €200+, with raw compute of an RX 6700 (~11.3 TFLOPS) — on a ~€50 board.


🛠️ Tools and method

ToolWhat it measures
vkpeakFP32/FP16/int8 throughput via Vulkan
clpeakmemory bandwidth and OpenCL throughput
sysbenchCPU stress/benchmark (also used by the Tuner)
Unigine Superposition / HeavenOpenGL graphics benchmarks
In-game MangoHudFPS and frametime in real games
custom telemetryclock/temp/power/fan via amdgpu sysfs, k10temp, nct6686

Sources