daytonb, you're the man, thanks  everything you said worked. Hence 

The stupid ram mistake purchase was my part not doing enough pre-buy study. No rudness taken, no worries.

BIOS was updated to 3.2 to begin with, may be this help too???

Give you a good weekend laugh, I was one of those who burnt the BIOS using windows_flash early batches. Well the full story, my 7 years old full AMD PC finally died early APRIL and hence the next day hastily went out to buy this all new Ryzen1700x. Taichi was a late minute decision because of availability, else I would have taken the KILLER version saves me around USD100, but was cleared out cause nobody wants Taichi the not high C/P buy. Anyway Corsair was also in the same shop, hence ram and psu, finally Intel600P was also not my choice, but you know now a hasty purchase.

The old HIS_R9_290x_Ice_Q was used until October2017 got the HIS_Vega_64. Called to the dealer every month to reserve, but he said only for regular customers and miners, which indirectly means not me, but after monthly enquiries he sold a unit to me. 

EDIT UPDATE 3rd July 2018, 2933mhz timing details

Edit 30th July 2018 have a better 2933mhz voltage and timing

Hi everyone, sharing my story and asking for timing understanding.

X370 Taishi bios 2.4 + Corsair 2x16GB dual rank dual sided = CMK32GX4M2B3000C15

Since can't run XMP = 3000Mhz 15-17-17-35 1.35V

Was able to run 2666Mhz using 14-13-13-32-51-1T undervolt to 1.2V stable since May 20th.

Later games start to crash, hence up the ram voltage to 1.26V on June 17th and stable again kept 2666Mhz using 14-13-13-32-51-1T.

Around late August games crash again, hence up the ram voltage again to 1.3v and stable till today. 

From calculating timing nanoseconds point of view, does it go like below example? or something else please demonstrate, thanks.

3000Mhz = 0.00000000033seconds each cycle and it needs to pass through = multiply by latency 15-17-17-35 = total 84 steps hence assumed to delay 2.8e-8seconds

while the 2666Mhz is 0.00000000038s multiply by 14-13-13-32 = 72 cycles = 2.7e-08seconds delay

which is 3.68% faster in latency but -11% less work done per second, so it's still slower than XMP. Something like that? Hence I'm still not making full use of my ram's original capability by -7.456%.

Everything else default, no overclock whatever.