Discussion in 'Videocards - AMD Radeon' started by OnnA, Oct 29, 2020.
Finally you were able to get one.
Enjoy the performance.
My every day gaming setting for the 6800XT is 2650MHz@1005mV with WattMan at 15% without any additional power setting using MPT.
Here are the Time Spy scores and the Port Royal Stress test for stability for these settings:
These settings run without issues. Played many sessions of Assassin's Creed Valhalla and Cyberpunk with Ultra in game setting using these.
In Assassin's Creed Valhalla clock hovers mostly around 2600MHz.
What I noticed is that there is a bug in the drivers. Sometimes when I open Assassin's Creed, the game clock displayed by Radeon overlay is about 50MHz lower i.e. in range of 2530-2550MHz.
If this happens then I need to close the game and start it again and then GPU runs in range of 2590-2620MHz.
Tried the trick with changing WattMan profile without needing to close the game. It does not work for me. I need to close and restart the game if GPU is clocking 50MHz too low.
The maximum I can reach with my GPU is 2750@1005mV with MPT set to 280W for GPU and 330A for TDC with 15% in WattMan. This is close to limit of 375W total available power.
Also tried 285W without any improvement.
Here is the Time Spy score with those settings using MPT:
I was never able to reach Time Spy score of 20K with any MPT settings.
It is always about 300 points too short.
are you running the steam version or have you downloaded a copy by itself, you could probly break 20k otherwise with just 3dmark running
@Turbogear I hit 20K with 2550Mhz real frequency with MPT set to 330W for GPU and 350A for TDC.
(it's not stable for games that use RT but for TimeSpy it was fine)
6900xt are easier to find then 6800xt's. They are not that much more expensive. Its like 1300-1400eur for 6900xt and 1100-1200eur for 6800xt.
thats the store I got my 6800xt.
I have Ubisoft Connect version of Assassin's Creed Valhalla.
Thanks for feedback.
I think important is to find a good setting which gives best possible performance and is stable in games.
In Radeon VII it was much easier to configure as it did not provide that much margin for undervolt tuning.
With 6800XT somehow I have the feeling that one can try almost endless combinations for Frequency and voltage.
I liked the Frequency vs Voltage curve in Wattman that was in previous version of Radeon GPUs.
They removed it for 6800XT.
i was talking about 3dmark, when you go to your profile on 3dmarks site you can download an offline version of 3dmark timespy where you should be able to get higher gpu/cpu scores for the benchmark
Sorry I misunderstood you.
That's enteresting, I did not know that this makes a difference.
I have the Steam version.
Seems i got lucky with my R6800 , it runs stable with driver max OC ( 2500-2600 gpu 2150 mem +15% power). You think there will be option to go above that ? Seems there is room for extra performance.
Stock temp 72C ( torture)
OC temp 77C .
Performance gain is around 10% depending on which game i test ( 10% in RDR2 and Cyberpunk).
Finally got the MPT to work and to actually set up values.
So on my 6800xt gaming x trio
Any value above :
or above the 15% power target
Results in values in drivers to be 500mhz on the clock and 0 powertarget
I can put the vram to 2600mhz without crashing but the clock goes to 450mhz
I can finally go over the 265w up to 300w max but results in nothing really, still caped ou at 2600mhz after that results in artifacts, it needs more voltage.
So max overclock is 2600mhz and 2150mhz the 9% powertarget. Anything above is just wasting energy on my card, unless I can flash a bios with more voltage.
Here is a good article at Igor's Lab about using MPT:
As far as I understood, it is not possible to increase the three parameters you mentioned above. Changing these will lead the GPU to go to Fail Safe state where clock is stuck at 500MHz.
Anybody tried the new driver 21.1.1?
I installed it and it seems the boost curves are behaving differently now.
To get same performance like before I need to go down in voltage from 1005mV to 980mV.
At 1005mV now the performance is worse.
So new setting I am using now is 2650MHz@980mV with Power limit in Wattman at 15%.
Below are the test result by using these new setting including Port Royal Stress Test:
By the way the memory clock bug is still not fixed.
When my ASUS XG27W monitor is set at 165Hz refresh rate then the memory clock does not decrease even when system is idle. Memory always run at full speed all time.
If I reduce refresh rate of monitor to 144Hz then the memory clock goes down at idle.
you should be able to download timespy if you wanna just get your max speed without steam and other programs running (https://www.3dmark.com/settings), i have pretty much accepted my memory clock is running at full all the time until they get hardware gpu scheduling implemented again, im running 1x1440p@144hz, and 1x2160p@60hz only if i set my screens to 1440p@60hz or lower i cant get the memory to downclock. it also doesnt look like your using MPT to set your voltage so your probly still hitting 1150mV running those clocks
I will try to download Time Spy as you hinted.
Yes limiting the voltage with MPT is the next point on my agenda. It always hit 1150mV.
I guess I need to experiment with MPT regarding what value should be set for max voltage. I need to see what is the lowest MPT max voltage setting were my GPU is still stable under load.
Did you adjust only Maximum Voltage GFX or did you also experiment with Max Voltage SoC?
Igor was mentioning on his website the SoC voltage does not have that large influence.
I have tried the driver and haven't checked the performance really....
IMHO was more stable in the division 2 which I have played yesterday, and they finally fixed the monitor bug not only for G9 but also C49RG90.
A keeper for me for sure!
P.S. The memory "bug" must not necessarily be a "bug" per se.... This is and was a more or less normal behavior for multi-mon and/or high refresh rate monitors....
I believe Spajdr posted a possible "solution" on modifying some values in CRU in the last driver thread.
IMHO I don't believe it will be fixed as it is somewhat.... expected?
it might be possible for them to fix the downclocking bug with infinity cache usage, though not sure about dwm being that limited, somehow ampre graphics cards dont have that issue with there vram (nv broadcast keeps the mem clk up).
More states for the VRAM could also help, isn't that why it works on Vega and Fury since the HBM uses less voltage and it can go into a mid-clock state not just idle and full speed.
Forgot the specifics of how it worked the VESA standard, timings and how it impacted things but short of it is basically to just use CRU (Think it got updated again a few days ago as well. ) and fiddle with the timings finding a combination that works around this.
True. HBM was in a different league in that area.... Would be nice to see a sub-step for the memory e.g. clocking at 400 MHz during such scenarios as high refresh and/or resolution displays etc...