Socket power and energy on Strix Point¶

Abstract¶

This study measures socket power and energy on a Framework 13 with an AMD Ryzen AI 9 HX 370 (Strix Point) running Gentoo Linux, for two workloads. The first is Whisper speech-to-text with the turbo model on the XDNA 2 NPU and on the Radeon 890M through Vulkan and ROCm. The second is llama.cpp prefill and decode on the Radeon 890M, comparing Vulkan with ROCm on four models. All measurements were taken on AC power during one night in September 2026.

For speech-to-text, the NPU path drew about 56% of the socket power of the Vulkan path but took about three times as long, so it used 1.5–1.7 times as much energy per minute of audio. For llama.cpp, socket power was close to 33 W on both backends, so energy per token followed speed. Vulkan was faster than ROCm in every test up to 4,096 prompt tokens. At 16,384 and 32,768 prompt tokens ROCm was 5–10% faster on the mixture-of-experts model only.

Socket power is the firmware-reported figure for the whole processor package. It is not wall power or battery drain.

Experimental methods¶

Platform and software¶

The measured Gentoo packages were app-accessibility/whisper-cpp, sci-misc/llama-cpp, sci-ml/fastflowlm and sci-ml/lemonade from the stuff overlay. The Vulkan runs used the installed packages. The ROCm runs used the same ebuilds compiled with HIP for gfx1150 and run from their build directories, without installation.

Power measurement¶

Power is average_socket_power from the amdgpu gpu_metrics interface, sampled every 0.5 s. Energy is the trapezoidal integral over each measured window, with linear interpolation at the window boundaries and no idle subtraction.

Before each cell a gate required AC power, a battery that was not discharging, a hottest thermal zone below 70 °C and a one-minute load average below 1.0. It then waited until every socket sample stayed below 8 W for 30 consecutive seconds. The idle readings recorded before the speech-to-text cells were 5.1–6.6 W.

Speech-to-text workload¶

The corpus, models and scoring are those of the Whisper on FLEURS study: 100 English and 100 Russian FLEURS recordings, 15.9 and 19.6 minutes of audio. Only the turbo model was measured, on three paths: whisper.cpp on Vulkan, whisper.cpp on ROCm, and FastFlowLM on the NPU. Each path and language was measured three times, for 18 cells. The power window of a cell is the transcription itself. The NPU model was loaded before the idle wait.

llama.cpp workload¶

These are the model files of the NPU versus Radeon 890M study. Their SHA-256 values are in the evidence notes; they were computed after the run.

Each model and backend was one llama-bench call with all layers on the GPU (-ngl 99) and flash attention on. A discarded 32-token prefill came first and absorbed model loading. The first run then measured prefill of 512 and 4,096 tokens and decode of 128 tokens with five repetitions. A second run measured prefill of 16,384 and 32,768 tokens with three repetitions. Backend order alternated per model.

Each test has its own power window, which includes llama-bench's warm-up for that test. Energy per token is the mean socket power of the window divided by llama-bench's mean tokens per second. Prefill and decode here are synthetic llama-bench tests, not served requests.

Results¶

Speech-to-text¶

Means of three repetitions.

The NPU path drew about 56% of the socket power of the Vulkan path and ran 2.8–3.0 times as long. It therefore used 1.5–1.7 times as much energy per minute of audio. About 1.3 W of its 18 W was reported in the NPU power domain; the remainder was the rest of the package.

ROCm used 10–20% more energy per audio minute than Vulkan at the same socket power, because it was slower. Energy per audio minute varied between repetitions by under 1% on Vulkan, under 2% on the NPU and up to 7% on ROCm.

The NPU path was also the least accurate. Its error rates are those of the Whisper on FLEURS study, where the NPU transcripts are analyzed.

llama.cpp speed¶

Mean tokens per second, listed as Vulkan / ROCm.

Vulkan was faster than ROCm in all twelve comparisons: by 16–23% in decode, by 7–32% in 4,096-token prefill and by 21–44% in 512-token prefill. The smallest difference was the 4,096-token prefill of Qwen3.6-35B-A3B, the mixture-of-experts model.

llama.cpp energy per token¶

Joules per token, listed as Vulkan / ROCm.

Mean socket power during decode was 33.0 W for every model and backend except Qwen3.6-35B-A3B on ROCm, at 31.4 W. During 4,096-token prefill it was 33.0–36.5 W. With power nearly equal, energy per token followed speed.

The 512-token prefill windows lasted 1.9–8.5 s. Their power readings, from 29.6 to 41.5 W, are dominated by the transition from the previous test, and no energy figure is reported for them.

Long prefill¶

Mean tokens per second and joules per token, listed as Vulkan / ROCm.

The two Qwen3.5 models kept a Vulkan advantage of 21–24% at both lengths. Gemma-4-E4B-it on ROCm fell behind as the prompt grew: Vulkan was 27% faster at 4,096 tokens, 53% at 16,384 and 94% at 32,768. The cause was not investigated.

Qwen3.6-35B-A3B reversed. ROCm was 5% faster at 16,384 tokens and 10% faster at 32,768, and used correspondingly less energy per token. Decode on the same model was still 22% faster on Vulkan.

These windows lasted 44–504 s. Socket power was 33.0–34.2 W in every long window except Gemma-4-E4B-it on ROCm at 32,768 tokens, at 31.3 W.

Discussion and limitations¶

Socket power covers the processor package: CPU cores, integrated GPU, NPU and their shared logic. It is not wall power and not battery drain, and no idle power was subtracted. A path that leaves the machine idle sooner is credited only for the measured window.

The near-constant 33 W under Radeon 890M load is consistent with a sustained package power limit. Throttle counters were not logged in this run, so the limit is inferred and not measured.

The speech-to-text comparison is between complete paths. The NPU path differs from whisper.cpp in runtime and model format (FastFlowLM q4nx against whisper.cpp q8_0) as well as in device. The result does not show that the NPU hardware is inefficient; it shows that this path used more energy for this workload. It also does not cover a case where the NPU runs alongside a GPU job, which was not measured.

The ROCm builds were uninstalled test builds of the overlay ebuilds on gfx1150 with ROCm 10.0.0. Other ROCm versions, other GPUs and tuned builds may behave differently. The llama.cpp tests measure throughput only and say nothing about answer quality.

The measurements come from one laptop on one night. Three or five repetitions are repeated measurements, not independent hardware samples.

Conclusion¶

On this Framework 13, Vulkan on the Radeon 890M was the fastest measured path for Whisper and for llama.cpp, and because socket power was nearly the same across Radeon 890M workloads it was also the one that used the least energy. ROCm was slower in every llama.cpp test except long prefill on the mixture-of-experts model. The NPU path for Whisper drew less power but used more energy per minute of audio.

Data availability and reproducibility¶

The speech-to-text table has one row per cell with WER, speed, mean socket power, socket energy, energy per audio minute, NPU and graphics domain power and the idle reading before the cell. The llama.cpp table has one row per model, backend and test with tokens per second, its standard deviation, window length, mean socket power and joules per token.

The sanitized energy evidence archive and its method and verification notes are pinned to signed commit d22bc59. The archive keeps the raw power records of every cell, the transcripts, the llama-bench output and the exact command of each cell. Verification rescores the speech-to-text cells, recomputes their energy from the raw samples, checks the llama.cpp rates against the llama-bench output and regenerates both tables. The llama.cpp windows were saved as durations only, so their start is rebuilt from the power samples; the recomputed mean power matches the recorded values to within 0.02%.

References and setup guides¶

• whisper.cpp
• llama.cpp
• FastFlowLM
• Ryzen AI NPU setup on Gentoo
• llama.cpp, Ollama and llama-swap setup
• GPU hardware identification