GPU Memory Benchmarks

Append one row per official warmed GPU-memory benchmark. Each scenario is prewarmed independently, then unused allocator blocks are cleared and PyTorch peak counters are reset without unloading model weights or cached clone prompts.

Process peak is PyTorch’s maximum allocated memory for the TTS process. Device peak is sampled through nvidia-smi and includes every process on the selected GPU. Official comparisons require the same image, settings, workload, and an otherwise idle GPU.

Scenario shortcuts: RV random voice, DV designed voice, CR uncached direct clone with transcript, and CC cached clone with transcript.

Run	GPU	Calls	RV process peak	RV device peak	DV process peak	DV device peak	CR process peak	CR device peak	CC process peak	CC device peak	Comment
06.10.2026 17:43:51 - 1.0-snapshot	NVIDIA GeForce RTX 5070 Ti	5	2116.9	4706	2116.9	4706	2118.9	4608	2115.2	4764	v1.0-snapshot baseline; Qwen3-ASR stopped; clean OmniVoice container; Windows desktop consumers present
28.10.2026 17:43:10 - 1.1-snapshot	NVIDIA GeForce RTX 5070 Ti	5	2107.6	4822	2116.7	4831	2120.1	4728	2115.1	4864	PR283 post-fix; source-mounted 3481d51 plus local inference guard; Torch2.10-cu128 Transformers5.17; remote ASR; Windows desktop present; warmed short-reference workload, not cold15s peak test