Are there any benchmarks comparing performance with CuDNN?
[0] http://i.imgur.com/twhTpcC.jpg
[1] http://i.imgur.com/1peXVnq.png
[3] tldr: 16 GByte HBM2, 25 TFlop FP16, ~1000-1500 dollar.
[4] https://www.newegg.com/Product/Product.aspx?Item=N82E1681410...
Using OpenCL code with batch size which favors one or the the other is enough to cause this (and much) higher delta.
DeepBench isn't a benchmark, it's a benchmarking tool overall there is very little chance that given the current state of NVIDIAs BLAST libraries and the rest of their eco system that Vega is going to be beating it's hardware.