Acoustic-variation scaling

Run 20261102T181141Z-14c926d8; NVIDIA GeForce RTX 5070 Ti; image sha256:80e6c58ff65124fa7f0276c8f3f4e4fb45b387e2f31259a90bab2582c15d910a. Matched two-minute English songs, fixed seeds per count, eight steps, resident models. Every count uses a fresh worker and a full-length excluded warmup.

Variations	Mean batch s	Mean s / rendition	Throughput vs one	Extra batch s / added rendition	Throughput gain over previous	Device peak MiB
1	23.55	23.55	1.00x	n/a	n/a	12591
2	28.17	14.09	1.67x	4.62	67.2%	12987
3	33.81	11.27	2.09x	5.63	25.0%	13537
4	40.62	10.16	2.32x	6.82	11.0%	14161

Stage means below come from rounded native log summaries, matched to measured job IDs. They are explanatory timings, not GPU kernel profiles. NAR includes acoustic chunk setup/solve; VAE sums all renditions, which decode separately. The residual includes rounding and unlogged work.

Variations	Score s	Semantic s	Acoustic NAR s	All VAE decodes s	Residual s
1	7.00	11.53	4.27	0.77	-0.02
2	6.70	11.13	8.80	1.50	0.04
3	6.70	11.30	13.40	2.32	0.08
4	7.23	11.97	18.20	3.11	0.11

Throughput is based on aggregate pipeline time divided by rendered track count. It excludes ASR, cold/warmup requests, encoding and HTTP delivery. Memory peaks are sampled whole-device observations including desktop/driver allocations. More variations share a composition and semantic stream, but have distinct waveform hashes.

Small differences need repetition before being treated as a stable hardware limit. These measurements do not establish memory requirements for five-minute songs, other prompts, or larger batches. ASR remains an imperfect comparative judge. Total original suite wall time, including warmups and ASR: 764.86 seconds. This offline analysis does not alter captured requests, audio, timings or the original run manifest.