GPU: NVIDIA A40 Python Version: 3.12.3 PyTorch Version: 2.8.0+cu128 CUDA Version: 12.8 Driver Version: 595.91.07 L2 Cache: 6.0 MB Theoretical Bandwidth: 696.0 GB/s WARNING: Logical bytes != DRAM transactions. Achieved GB/s > theoretical DRAM ceiling indicates coalescing reuse. The theoretical ceiling applies to real traffic. Quattra Tiled Benchmark (3 rounds of 20 iters per L, median & p95 reported) L N Mod_med Mod_p95 Cond_med Cond_p95 C16_med C16_p95 Tile_med Tile_p95 Speedup Naive GB/s Tiled GB/s Max Abs Diff --------------------------------------------------------------------------------------------------------------------------------------------------------- 16 256 0.00300 0.00321 0.00304 0.00316 0.00304 0.00310 0.00304 0.00314 0.998 1.69 1.68 4.77e-07 32 1024 0.00305 0.00324 0.00304 0.00321 0.00303 0.00313 0.00312 0.00320 0.972 6.75 6.56 4.77e-07 48 2304 0.00305 0.00321 0.00298 0.00311 0.00297 0.00308 0.00317 0.00319 0.939 15.50 14.55 4.77e-07 64 4096 0.00309 0.00313 0.00297 0.00310 0.00300 0.00309 0.00317 0.00321 0.937 27.55 25.81 4.77e-07 128 16384 0.00345 0.00349 0.00312 0.00319 0.00297 0.00313 0.00323 0.00326 0.921 110.34 101.59 4.77e-07 256 65536 0.00568 0.00571 0.00525 0.00529 0.00404 0.00406 0.00401 0.00406 1.007 324.76 326.90 9.54e-07 512 262144 0.01452 0.01473 0.01388 0.01393 0.00853 0.00859 0.00753 0.00756 1.133 614.44 696.45 9.54e-07 1024 1048576 0.05275 0.05291 0.05415 0.05419 0.02755 0.02764 0.02587 0.02593 1.065 761.25 810.66 9.54e-07 2048 4194304 0.22271 0.22333 0.20684 0.20698 0.10493 0.10503 0.10019 0.10025 1.047 799.46 837.27 9.54e-07 4096 16777216 0.88426 0.88467 0.82875 0.85573 0.52296 0.52356 0.48240 0.48282 1.084 641.62 695.57 9.54e-07 8192 67108864 4.47661 4.48817 4.43037 4.43314 2.31546 2.31684 2.06847 2.07124 1.119 579.66 648.87 9.54e-07 Rule 3 Distribution Audit (Per-Round Separation Verification): L Tiled Rounds (ms) Slowest Tiled Fastest Control Separation Pass ------------------------------------------------------------------------------------- 2048 [0.10016, 0.10019, 0.10020] 0.10037 0.10185 PASS 4096 [0.48089, 0.48240, 0.48253] 0.48294 0.52194 PASS 8192 [2.06829, 2.06847, 2.06854] 2.07172 2.31300 PASS Pre-Registered Pass/Fail Evaluation: Pre-Registered Rules: (1) >3% speedup vs best control, (2) wins at >=2 of {2048, 4096, 8192}, (3) slowest-of-3 tiled < fastest-of-3 control. RESULT: PASS — skew_tiled achieved >= 3% speedup with statistically separating bounds at 3/3 large sizes.