AMD Memory Tweak - Read/Modify Timings on the fly!

Ι haven't check for memory errors ,just gaming and unigine valley loop for about an hour. Any suggestion for such a program to test for memory errors?

 

Asus. Default bios. It crossed my mind to update the bios with LC but i hesitate

 

Lc bios stopped work correctly after 19.1.2 driver so you will lose your time with this. I use default bios too the locked one.

I have to raise soc voltage above 950mv if I clock hbm frequency above 1015mhz. Give a try with 1000mhz or lower and 950mv.

 

I find hwinfo pretty accurate. Let it run while you do benchmarking and gaming. Even one error should make the card unstable. Let us know how it went.

 

Υou are talking about aida gpu stress test? Ok I will check it when I go home.

 

No sorry i didnt meant aida but hwinfo. Fixed it.

https://i.**********/QC9XTSDL/hyHiteA.png

 

A good memory test that also stresses the card is OCCT.





@Undying Vega doesn't have memory error reporting with HWinfo but it's good for Polaris.

 

Ok I will rub occt and I 'I'm report as soon as I can.

 

I didnt know Vega didnt have an option in hwinfo, thanks.

 

I think it's do do with error correction technology being present on GDDR5 so hwinfo can read the data where as HBM doesn't seem to have it. The Artifacts are easily noticable on HBM where as my old Nvidia 1070 for example would keep correcting memory errors as the frequency raised and artifacts were less obvious so only the drop in performance caused by the error correction or the hwinfo report would give it away.

 

I did. Even at stock 945Mhz, voltage must be equal or more than 951mV otherwise it stays at 800Mhz. It must be Asus way of things

 

It's a reference design or strix ? Verry strange. If it is a reference one you can try to flash the unlocked bios with another's brand stock reference bios and see how it goes.

 

It's Strix

 

Testing tCL because I don't think it's working. Tests using Superpostion because it has very consistant results and is susceptible to memory timings.

tCL 1080p/Extreme over 2 runs

cl 18 4522/4524 Average 4523
cl 16 4526/4523 Average 4524.5
cl 14 4523/4522 Average 4522.5
cl 12 4524/4522 Average 4523

Conclusion - Limited data reults but I would say it's not working or values below 16 are ignored.

 

Guys...Win10 have a very specific and "curious" behavior towards TimeSpy's combined score if you have Ryzen R7 (don't know about R5) cpus.

1703, 1803, normal channels don't use the full processor's potential. If you run AMD overlay, you'll see that while on cpu test there is >80% cpu usage (even more, usually), on combined test, it hovers around 50% (sometimes a bit more, sometimes a bit less). Combined test relies on cpu to feed the gpu, so the gpu also stays around 70-80%.

Insider channels (at least for me), show a mixed behavior... Sometimes there is a sudden bump in performance on combined (around 80% cpu usage), sometimes it's the "normal" 50%. On 1903 Insider Slow (and fast), there is proper (almost) cpu utilisation, so you're noticing elevated scores in combined test (higher cpu +gpu usage, higher fps).

It has nothing to do with Radeon Settings driver version etc...it's a Windows10 "problem".

Needless to say, i79900K (and every Intel cpu) had no such problems from the very first day it was released...full utilisation on cpu test, full on combined test.

*exact percentages may (and will) vary...but the problem's core is as I described it.

 

Ahha, strix' bios is a little bit strange.
So you can not flash any other bios but you can try the latest bios from Asus page.

 

@Dekaohtoura
So probably improvements in the windows scheduler in 1903 which determines if an application is using a real core or a virtual core. Intel already had good support for their hyperthreads so not much improvements to have. AMD however should be getting a nice boost. It's not a problem though and the combined results are consistent on 1903.

 

Holy scores, Batman!

Can't wait for these drivers to hit public release!

Now, regarding HBM tweaks, you're referring to post #42, right?

 

Hmm, seems like CL14 is no go (CL16 is the Best for my setup)
Tweak: CL16 up to CL22 Working (Vega64 HBM2 Samsung)

Here:

Spoiler: Test

 

Ι tried cl16 too and really seems better than cl12 and cl14 ,I score one fps more on sottr, 106fps instead of 105fps with cl12/cl14.
To summarise with stock timings it scores 101fps in sottr with your timings at post #42 I had 104fps, changing cl12/trcd12/tcwl6/twrtl8/twr12/tcke1 it scores 105fps and with the same timings with cl16 it scores 106fps. 5 fps from timings not bad at all.