| { |
| "schema_version": 1, |
| "date": "2026-07-12", |
| "hardware": { |
| "gpu": "NVIDIA GeForce RTX 5060 Ti", |
| "compute_capability": "SM120", |
| "user_batch_size": 1 |
| }, |
| "software": { |
| "python": "3.12", |
| "torch": "2.7.1+cu128", |
| "torch_cuda": "12.8", |
| "flashinfer": "0.6.14", |
| "nvidia_cutlass_dsl": "4.5.2" |
| }, |
| "natural_eight_text_suite": { |
| "definition": "Pipeline and T3 seconds are sums of per-case medians; aggregate RTF is summed pipeline time divided by summed generated-audio duration; T3 throughput uses valid speech tokens; memory is maximum per-case median PyTorch CUDA peak.", |
| "d16_quality": { |
| "bf16": { |
| "audio_seconds": 98.64, |
| "pipeline_seconds": 9.304738, |
| "aggregate_rtf": 0.0943303, |
| "t3_seconds": 7.512483, |
| "valid_tokens": 2466, |
| "t3_tokens_per_second": 328.254, |
| "peak_allocated_bytes": 6423756800, |
| "peak_reserved_bytes": 7107248128 |
| }, |
| "nvfp4_w4a16": { |
| "audio_seconds": 92.52, |
| "pipeline_seconds": 8.072466, |
| "aggregate_rtf": 0.087251, |
| "t3_seconds": 6.330483, |
| "valid_tokens": 2313, |
| "t3_tokens_per_second": 365.375, |
| "peak_allocated_bytes": 5214659584, |
| "peak_reserved_bytes": 6322913280 |
| }, |
| "nvfp4_vs_bf16": { |
| "aggregate_rtf_speedup": 1.081136, |
| "t3_throughput_speedup": 1.113087, |
| "peak_allocated_reduction_bytes": 1209097216, |
| "peak_allocated_reduction_percent": 18.8223, |
| "peak_reserved_reduction_bytes": 784334848, |
| "peak_reserved_reduction_percent": 11.0357 |
| } |
| }, |
| "d24_fast": { |
| "bf16": { |
| "audio_seconds": 93.36, |
| "pipeline_seconds": 6.797084, |
| "aggregate_rtf": 0.0728051, |
| "t3_seconds": 5.04703, |
| "valid_tokens": 2334, |
| "t3_tokens_per_second": 462.45, |
| "peak_allocated_bytes": 6539530752, |
| "peak_reserved_bytes": 7411335168 |
| }, |
| "nvfp4_w4a16": { |
| "audio_seconds": 91.8, |
| "pipeline_seconds": 6.483264, |
| "aggregate_rtf": 0.0706238, |
| "t3_seconds": 4.748782, |
| "valid_tokens": 2295, |
| "t3_tokens_per_second": 483.282, |
| "peak_allocated_bytes": 5129031168, |
| "peak_reserved_bytes": 6064963584 |
| }, |
| "nvfp4_vs_bf16": { |
| "aggregate_rtf_speedup": 1.030886, |
| "t3_throughput_speedup": 1.045046, |
| "peak_allocated_reduction_bytes": 1410499584, |
| "peak_allocated_reduction_percent": 21.5688, |
| "peak_reserved_reduction_bytes": 1346371584, |
| "peak_reserved_reduction_percent": 18.1664 |
| } |
| } |
| }, |
| "fixed_capped_work": { |
| "definition": "One fixed-seed case, five-repeat medians, 1024-token reservation. D16 caps at 32 speech tokens; D24 caps at 48, so compare quantized and BF16 only within each block size.", |
| "d16": { |
| "bf16_t3_ms": 88.363, |
| "nvfp4_t3_ms": 78.475, |
| "t3_speedup": 1.125998, |
| "bf16_pipeline_ms": 162.897, |
| "nvfp4_pipeline_ms": 153.235, |
| "pipeline_speedup": 1.063052, |
| "bf16_peak_allocated_bytes": 4165647360, |
| "nvfp4_peak_allocated_bytes": 2745680896, |
| "peak_allocated_reduction_percent": 34.0875, |
| "peak_reserved_reduction_percent": 30.0444 |
| }, |
| "d24": { |
| "bf16_t3_ms": 91.192, |
| "nvfp4_t3_ms": 85.574, |
| "t3_speedup": 1.065645, |
| "bf16_pipeline_ms": 180.938, |
| "nvfp4_pipeline_ms": 176.006, |
| "pipeline_speedup": 1.028023, |
| "bf16_peak_allocated_bytes": 4173167104, |
| "nvfp4_peak_allocated_bytes": 2754244096, |
| "peak_allocated_reduction_percent": 34.0011, |
| "peak_reserved_reduction_percent": 30.0689 |
| } |
| }, |
| "automated_quality": { |
| "warning": "Small eight-text, one-reference automated evaluation; not human MOS or a multi-speaker benchmark.", |
| "d16": { |
| "bf16_mean_score": 94.125, |
| "nvfp4_mean_score": 94.625, |
| "exact_transcripts_each": "8/8", |
| "hard_failures_each": "0/8", |
| "order_controlled_pairwise": "7 ties, 1 BF16 win" |
| }, |
| "d24": { |
| "bf16_mean_score": 95.0, |
| "nvfp4_mean_score": 93.125, |
| "exact_transcripts_each": "8/8", |
| "hard_failures_each": "0/8", |
| "order_controlled_pairwise": "6 ties, 1 BF16 win, 1 NVFP4 win" |
| } |
| }, |
| "source_free_load_diagnostic": { |
| "fresh_process_load_seconds_with_base_hash_verification_and_jit_cache_hit": 2.675073, |
| "allocated_bytes_after_load": 1444340224, |
| "reserved_bytes_after_load": 1478492160, |
| "prepared_projection_bytes": 283116000, |
| "prepared_projection_count": 120, |
| "short_generation_peak_allocated_bytes": 2548431360, |
| "first_same_process_generation_seconds_with_exact_shape_config_cache": 10.567298, |
| "second_same_process_generation_seconds": 0.153511, |
| "first_to_warm_ratio": 68.8376, |
| "notes": "Canonical weights were prepared without BF16 projection materialization or nvfp4_quantize. The first-generation measurement includes engine creation, CUDA graph capture, and per-process FlashInfer setup even though the exact shape's selected config was cached. A completely missing exact-shape tuning entry can take longer. Caches are GPU-specific and are intentionally not distributed." |
| }, |
| "interpretation_warnings": [ |
| "Do not present raw natural-suite wall-time ratios as pure NVFP4 speedups because BF16 and NVFP4 sampled different token counts.", |
| "PyTorch allocated/reserved peaks are not total board VRAM or nvidia-smi process usage.", |
| "The D24 versus D16 speed difference is primarily the block-size change, not quantization alone.", |
| "The D16 BF16 natural control had one token-cap hit and one duration warning; NVFP4 did not.", |
| "D32 was rejected for release use after cap hits and an OOM in the natural suite." |
| ] |
| } |
|
|
