cpu_memory
CPU-only Docker memory-limit benchmark for OpenAI-compatible speech requests.
The three historical rows ending with 28.10.2026 18:03:34 used FP32 CPU TTS. Later snapshots default to FP16 with optional FP32; check each new run’s per-attempt precision in DETAILS.md/runs.json before comparing memory requirements. The main server’s System toggle does not configure the separate CPU benchmark containers.
Each run starts one CPU container per scenario and Docker memory limit, then records the lowest passing limit as a conservative whole-GB RAM recommendation. The recommendation adds 512 MiB headroom to the lowest passing limit before rounding up to whole GB. The detailed pass/fail attempts are stored in runs.json and summarized in DETAILS.md.
The default task starts at 1536 MiB and uses practical steps up to 12 GiB; it intentionally skips obviously unusable sub-GB limits so the benchmark finishes faster.
The benchmark text is intentionally one realistic short request, around 100-200 characters. Results are not a guarantee for long text, concurrent requests, larger outputs, or different host memory behavior.
Scenario shortcuts:
RV: Random/no-prompt voice: no reference audio, no design instructions, no stored profile.DV: Voice design/generate voice: explicit design instructions, no reference audio.CR-NT: Direct clone reference without transcript. This may lazy-load ASR.CR-TX: Direct clone reference with transcript.SV-NT: Stored voice profile without transcript. This may lazy-load ASR.SV-TX: Stored voice profile with transcript.
| Date | Text chars | Version | RV | DV | CR-NT | CR-TX | SV-NT | SV-TX |
|---|---|---|---|---|---|---|---|---|
| 02.07.2026 21:29:38 - 0.2.1-snapshot | 140 | 0.2.1-snapshot | 2 GB | 2 GB | 6 GB | 2 GB | 7 GB | 3 GB |
| 06.10.2026 18:00:55 - 1.0-snapshot | 140 | 1.0-snapshot | 2 GB | 2 GB | 6 GB | 2 GB | 7 GB | 3 GB |
| 28.10.2026 18:03:34 - 1.1-snapshot | 140 | 1.1-snapshot | 2 GB | 2 GB | 6 GB | 2 GB | 6 GB | 2 GB |