GPU: NVIDIA A40 PyTorch Version: 2.8.0+cu128 CUDA Version: 12.8 Driver Version: 595.91.07 L2 Cache: 6.0 MB Theoretical Bandwidth: 0.0 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 0.00299 0.00319 0.00302 0.00316 0.00304 0.00334 0.00298 0.00328 1.012 1.70 1.72 4.77e-07 32 1024 0.00301 0.00362 0.00295 0.00311 0.00298 0.00310 0.00310 0.00312 0.951 6.95 6.61 4.77e-07 48 2304 0.00305 0.00321 0.00294 0.00312 0.00298 0.00315 0.00314 0.00327 0.934 15.69 14.66 4.77e-07 64 4096 0.00307 0.00329 0.00297 0.00309 0.00297 0.00305 0.00315 0.00319 0.943 27.62 26.04 4.77e-07 128 16384 0.00344 0.00355 0.00308 0.00315 0.00296 0.00309 0.00319 0.00322 0.926 110.88 102.66 4.77e-07 256 65536 0.00565 0.00578 0.00520 0.00525 0.00402 0.00406 0.00398 0.00401 1.010 326.11 329.26 9.54e-07 512 262144 0.01452 0.01463 0.01386 0.01393 0.00851 0.00858 0.00750 0.00755 1.134 616.29 698.83 9.54e-07