{"as_of":"2026-11-16","project_snapshot":"https://github.com/Casxt/fasth3-rtx50/tree/e3ab0ee","sources":{"full_generation":"https://github.com/Casxt/fasth3-rtx50/blob/e3ab0ee/benchmarks/linux-validation.json","transfer":"https://github.com/Casxt/fasth3-rtx50/blob/e3ab0ee/benchmarks/transfer-baseline.json","ab":"Selected numeric fields from retained project A/B summaries. Independent experiments, not a cumulative speedup series."},"units":{"seconds":"s","memory":"GiB = 2**30 bytes","bandwidth":"GB/s = 10**9 bytes/s"},"capacity_calculation":{"rows":226195,"width":5376,"mxfp8_group":32,"bf16_output_gib":2.265021800994873,"mxfp8_output_gib":1.1679018661379814,"residual_and_modulation_gib":4.530043601989746,"weight_arena_bytes":474486272,"weight_arena_gib":0.44189977645874023,"major_buffers_subtotal_gib":6.139845244586468,"head_dimension":128,"total_heads":42,"heads_per_group":4,"qkv_all_heads_payload_gib":6.795065402984619,"qkv_four_heads_payload_gib":0.6471490859985352,"reference_rgb_example":{"frames":360,"width":1344,"height":768,"channels":3,"uint8_payload_gib":1.0382080078125,"fp32_payload_gib":4.15283203125},"scope":"Capacity calculations only. Excludes padding, temporary layout copies, routing, library workspaces and other tensors; not allocator or whole-board peak."},"resident_output_ab":{"conditions":"15s reference + 15s requested output, 768p, one IP image (short edge 512), real first denoise step.","qkv_chunk_tokens":2048,"input_quantization_cache_bytes":0,"allocator_cap_gib":12,"bitwise_identical":true,"measurement":"Two unprofiled interleaved repeats per variant; CUDA activity statistics from separate profiled replays.","variants":{"host":{"mean_s":55.05533304999881,"allocated_peak_gib":6.9337334632873535,"cuda_activity_pct":98.27671309637587,"trace_idle_ms":955.0332529296875},"resident":{"mean_s":53.18976715000099,"allocated_peak_gib":8.490936279296875,"cuda_activity_pct":99.61473137565784,"trace_idle_ms":204.66173095703124}}},"weight_arena_ab":{"conditions":"Real 15s reference + 15s requested output R2VA, 768p, 226195 tokens; warm block0; measure blocks1-10 in first denoise step.","board_target_gib":11.5,"bitwise_identical":true,"measurement":"Main interleaved experiment only; independent strict recheck excluded.","variants":{"dynamic":{"mean_seconds":53.19708610000089,"allocated_peak_gib":8.490367889404297,"reserved_peak_gib":9.06640625,"measured_board_peak_gib":11.3466796875},"fixed1":{"mean_seconds":53.27623085000232,"allocated_peak_gib":8.490367889404297,"reserved_peak_gib":8.623046875,"measured_board_peak_gib":10.90234375},"fixed2-relaxed":{"mean_seconds":52.9575341499949,"allocated_peak_gib":8.932267665863037,"reserved_peak_gib":9.06640625,"measured_board_peak_gib":11.349609375}}},"full_r2va":[{"input_video_s":5,"output_s":5,"additional_images":1,"image_short_edge":512,"canvas":[1344,768],"fps":24,"transformer_forwards":8,"windows_5080_s":496.331,"linux_5090_s":286.746,"windows_process_working_set_private_commit_gib":[17.41,28.92],"linux_process_rss_gib":17.52,"windows_board_sampled_peak_gib":9.41,"linux_board_sampled_peak_gib":7.08},{"input_video_s":10,"output_s":10,"additional_images":1,"image_short_edge":512,"canvas":[1344,768],"fps":24,"transformer_forwards":8,"windows_5080_s":1344.252,"linux_5090_s":739.713,"windows_process_working_set_private_commit_gib":[17.9,29.8],"linux_process_rss_gib":17.92,"windows_board_sampled_peak_gib":10.69,"linux_board_sampled_peak_gib":9.57},{"input_video_s":15,"output_s":15,"additional_images":1,"image_short_edge":512,"canvas":[1344,768],"fps":24,"transformer_forwards":8,"windows_5080_s":2258.036,"linux_5090_s":1260.393,"windows_process_working_set_private_commit_gib":[18.34,30.64],"linux_process_rss_gib":18.1,"windows_board_sampled_peak_gib":11.35,"linux_board_sampled_peak_gib":9.88}],"pinned_256_mib_transfer":{"warmup":3,"repeats":20,"statistic":"Median CPU wall time including submission and completion wait.","pcie_under_load":"Gen5 x16 on both machines","windows_5080":{"h2d_gb_s":56.44,"d2h_gb_s":56.89},"linux_5090":{"h2d_gb_s":56.32,"d2h_gb_s":57.15}}}