This guide is supposed to show how the performance of the APU changes on various power levels and to help you pick the optimal power mode for your Strix Halo system.
All tests were run at fixed fans speed of 80% to make sure that cooling level is constant, CPU temp never exceeded 75 °C. For GPU becnhmarks please note that the GPU was in a passthrough mode to the Windows VM and only 12 CPU cores were passed as well, so there is at least 5% of a performance drop. In general, look at the percentage changes (relative to the base 55W result).
Everything single-threaded is omitted here, since a single core could never use the full power limit even on the lowest 55W value.
CPU - PassMark
Details:
- PassMark PerformanceTest for Linux, version 11.0.1002
-
Average of 5 medium tests for CPU, to lower the influence of the initial power boost (
pt_linux_x64 -i 5 -d 2)
| Metric | 55W | 85W | 120W |
|---|---|---|---|
| CPU Mark | 49 635 | 58 042 +16.9% | 62 805 +26.5% |
| Integer Math (MOps/s) | 193 409 | 222 270 +14.9% | 233 483 +20.7% |
| Floating Point Math (MOps/s) | 107 679 | 127 114 +18.0% | 138 255 +28.4% |
| Find Prime Numbers (MPrimes/s) | 263 | 286 +8.7% | 302 +14.8% |
| Random String Sorting (kStrings/s) | 71 735 | 88 253 +23.0% | 98 842 +37.8% |
| Data Encryption (MB/s) | 33 582 | 40 310 +20.0% | 44 246 +31.8% |
| Data Compression (KB/s) | 539 857 | 683 287 +26.6% | 777 048 +43.9% |
| Physics (Frames/s) | 4 802 | 5 477 +14.1% | 5 806 +20.9% |
| Extended Instructions (MMatrices/s) | 38 464 | 49 404 +28.4% | 58 761 +52.8% |
| * all percentage values are relative to the base 55W result |
CPU - Various Benchmarks
Details:
- sysbench - 1.0.20,
sysbench cpu run --threads=32 --time=30 - cpubench1a - 5.0,
cpubench1a -bench -nb 3 -duration 30 - 7-Zip - 16.02,
7z b
| Benchmark (Metric) | 55W | 85W | 120W |
|---|---|---|---|
| sysbench (Events/s) | 85 160 | 96 139 +12.9% | 99 218 +16.5% |
| cpubench1a (MT Score) | 3 483 | 4 288 +23.1% | 4 790 +37.5% |
| 7-Zip (Compressing MIPS) | 120 759 | 135 008 +11.8% | 140 471 +16.3% |
| 7-Zip (Decompressing MIPS) | 103 110 | 124 170 +20.4% | 135 859 +31.8% |
| * all percentage values are relative to the base 55W result |
GPU - Various Benchmarks
Details:
- UNIGINE Superposition - DirectX, 1080p Extreme preset
- 3DMARK Steel Nomad Light - Vulkan, 1080p
- 3DMARK Time Spy - 1080p
| Benchmark | 55W | 85W | 120W |
|---|---|---|---|
| PassMark 2D Mark | 621 | 706 +13.7% | 758 +22.1% |
| PassMark DirectX 10 | 59 | 65 +10.2% | 78 +32.2% |
| PassMark DirectX 11 | 112 | 134 +19.6% | 148 +32.1% |
| PassMark DirectX 12 | 63 | 67 +6.3% | 80 +27.0% |
| PassMark GPU Compute | 6 229 | 6 777 +8.8% | 7 311 +17.4% |
| UNIGINE Superposition | 5 247 | 6 103 +16.3% | 6 522 +24.3% |
| 3DMARK Steel Nomad Light | 8 494 | 9 883 +16.4% | 10 608 +24.9% |
| 3DMARK Time Spy | 8 891 | 10 230 +15.1% | 10 710 +20.5% |
| * all percentage values are relative to the base 55W result |
GPU - Gaming
Details:
- all games use pre-defined presets and built-in benchmarks
- all values are average FPS
- for Metro Exodus ray tracing was additionally set to
normal - for GTA5 the value is an average FPS of 5 different benchmarks
| Game | 55W | 85W | 120W |
|---|---|---|---|
| Cyberpunk 2077 (1080p low) | 61 | 70 +14.8% | 73 +19.7% |
| Cyberpunk 2077 (1080p high) | 53 | 62 +17.0% | 66 +24.5% |
| The Talos Principle (1080p ultra) | 102 | 113 +10.8% | 117 +14.7% |
| Metro Exodus Enhanced Edition (1080p ultra) | 51 | 59 +15.7% | 63 +23.5% |
| Red Dead Redemption 2 (1080p high) | 55 | 58 +5.5% | 66 +20.0% |
| Grand Theft Auto V Enchanced (1080p very high) | 105 | 118 +12.4% | 137 +30.5% |
| Black Myth: Wukong (1080p high) | 71 | 78 +9.9% | 84 +18.3% |
| * all percentage values are relative to the base 55W result |
GPU - LLM
Details:
- koboldcpp, Vulkan, full GPU offloading
- Q5_K_M
| Model (Metric) | 55W | 85W | 120W |
|---|---|---|---|
| Gemma 3 27B (Prompt Processing, Tokens/s) | 77 | 89 +15.6% | 95 +23.4% |
| Gemma 3 27B (Generation Speed, Tokens/s) | 6 | 6 +0.0% | 6 +0.0% |
| Qwen3 30B A3B (Prompt Processing, Tokens/s) | 92 | 95 +3.3% | 95 +3.3% |
| Qwen3 30B A3B (Generation Speed, Tokens/s) | 25 | 29 +16.0% | 29 +16.0% |
| * all percentage values are relative to the base 55W result |
Conclusions
Let's look on the average increase between different modes for each category:
| Category | 55W -> 85W | 85W -> 120W | 55W -> 120W |
|---|---|---|---|
| CPU - PassMark | +19.0% | +11.8% | +30.8% |
| CPU - Various Benchmarks | +17.0% | +8.5% | +25.5% |
| GPU - Various Benchmarks | +13.3% | +11.7% | +25.0% |
| GPU - Gaming | +12.3% | +9.3% | +21.6% |
| GPU - LLM | +8.7% | +2.0% | +10.7% |
| * for example, in case of Passmark we have a 19% performance increase when going from 55W to 85W budget and additional 11.8% increase relative to the original 55W performance when going from 85W to 120W, or a 30.8% in total |
Or represented on the chart:
A classic situation of diminishing returns, where 85W mode seems to be the sweet spot and a good compromise between performance gain and power consumption.
The 55W mode is suitable if you don't fully utilize your system resources (especially when you don't leverage all cores) or aren't concerned about performance loss. With that said, 55W limit is absolutely enough for the APU in general.
The 120W mode generally doesn't make much sense unless you aim to extract the absolute maximum from your system, with one exception - certain games respond more positively to this increase. My theory is that CPU-heavy games might slightly starve the GPU, so increasing the overall power budget alleviates this issue. Additionally, if you're just shy of reaching 60 fps at 1080p (a reasonable target for this system), this extra power might provide the necessary boost.
For LLMs, memory bandwidth (which APU's power limit doesn't affect) appears to play a more significant role than raw computing power, so again, exceeding 85W offers diminishing benefits.
In summary:
- 55W for light loads (background services, office work, multimedia)
- 120W for intensive gaming sessions
- 85W for mixed workloads and everything else
There is also a way to control power limits more gradually, read more here.
