Problem was in memory disposition.
I had 128 Gb of memory in first config. After I got 3970 x I took out 64GB and put it into new config.
Remaining RAM modules were put in first four slots. So only two channels were used. Memory clocks were set to default / auto.
In previous configs setting to auto meant that would work at declared clocks, but in both AMD configs that meant clocks were set to 2133 MHz even tough memory was 3000 GHz.
After I put modules one to each four channels and clock was set to 3000 GHz my vray benchmark result was 27000 and mentioned scene render time was 35 mins.
Which is huge difference compared to first result but difference is still 1,6 x (vray benchmark) and 1,8-1,9x (evermotion scene) slower compared to 3970x.