OWLSD-Turbo CPU image generation on the GoS1 bench: 0.21 Wh per image at 8 inference steps, 512×512
SD-Turbo on Ryzen 9 7900 CPU · 8 steps · 512×512 · 0.21 Wh / image · ΔW ≈ 30 W during the run
SCOPE: Device layer only (GoS1: AMD Ryzen 9 7900). CPU diffusion path; ROCm not in use for this measurement. Network and CPE excluded. No amortised training cost.
OWL Finding: SD-Turbo CPU image generation on the GoS1 bench: 0.21 Wh per image at 8 inference steps, 512×512
measured 2026-04-29
https://wattlab.greeningofstreaming.org/findings/sd-turbo-cpu-image-first-run
Greening of Streaming — wattlab.greeningofstreaming.org
Source measurement
Caveats
- n=1. This is the earliest stored SD-Turbo CPU result on the GoS1 bench (2026-04-29).
- The CLAUDE.md prose at the time of import cites the headline as 0.2063 Wh; the stored result file reports `delta_e_wh = 0.2099 Wh`. Numbers here are taken verbatim from the result file.
- Step count (8) was chosen to extend wall-time enough for reliable P110 polling — it is above SD-Turbo's native 1–4 step operating point. The energy reading is therefore higher than a minimum-step generation would produce.
- CPU path uses fp32 weights (PyTorch ROCm is not in use for the CPU diffusion code path). A GPU comparison on the same model lives in the broader image-generation workstream.
What was measured
One image generated by stabilityai/sd-turbo via the diffusers library on CPU. 8 inference steps, 512×512 output. Energy measured at the wall via a Tapo P110 plug at 1 Hz polling, with the same focus-mode + lock-file protocol every OWL measurement uses.
Numbers, from the stored result
delta_e_wh: 0.2099 Wh- Inference steps: 8
- Output size: 512 × 512
- Confidence: 🟢 under the CR-028 Phase 2 CI model
- Backend: AMD Ryzen 9 7900, CPU path
What this measurement does not establish
- It does not compare CPU and GPU energy on the same model. SDXL-Turbo and the GPU paths are separate stored runs.
- It does not measure image quality. SD-Turbo's native 1–4 step output is a separate question.
- It does not generalise across content. Single prompt, single run — variance across prompts is not characterised here.
Methodology → (docs/wattlab_traffic_light_confidence.md)
imagesd-turbocpudiffusion