Question XFX RX 580 GTS underclocking as soon as it starts going to full power, need help

CatPlayer

Commendable
May 12, 2016
41
0
1,530
0
XFX RX 580 GTS XXX Edition underclocking when going to 100% usage, as soon as it goes below 90% its clock goes to the original speed of 1366, makes no sense to me.
Graph picture from wattman
This happens with any game that is pushing it to its limits or furmark, in fact, the spike you see there was a furmark test on the 1440p preset. Temperature does not seem to be an issue as you see it didnt hit 80 degrees any time and the underclock happens as soon as the benchmark starts which makes no sense.
I've tried undervolting and fan curve but none of this make a difference, which further proved it may not be a temperature issue

Specs:
Motherboard: MSI B250M PRO-VD
CPU: Intel Pentium G4560 @3.50GHz
Memory: 2x4GB DDR4 @2133Mhz
GPU: XFX RX 580 GTS XXX 8GB GDDR5
VBIOS: 113-58085STD1-W90
Driver: Radeon Software Version 19.2.3
OS: Windows 10 x64 1809
 

CatPlayer

Commendable
May 12, 2016
41
0
1,530
0
System Requirements:
Minimum Power Supply Requirement: 500W
XFX Recommended Power Supply: XFX 550W
Power Connector: 8-Pin
EVGA 500B1 is my power supply, should be more than capable of handling this GPU, even more considering I got a very low consumption MOBO and CPU
 

tennis2

Notable
Nov 12, 2018
1,192
45
1,140
73
You ought to be able to get 1366MHz @ 1100mV. The GloFo 14nm process freq/voltage curve has an inflection point at ~950mV. Anything below that is going to be a flatter curve than anything above.
Also, 925mV for 2000MHz VRAM should work (I can do 900mV).
The VRAM voltage is going to act as the lowest load voltage for any of the core P-states (it will still idle lower, ~800mV IIRC). ie, if you have VRAM @ 925MHz but core @ 1100MHz/900mV, you'll see that it only goes down to 1100MHz/925mV.

Turn "Frequency" to "Dynamic" and "Voltage" to "Manual"
If you post a screenshot or list the 7 core frequencies, I can generate a voltage curve to get you in the ballpark.
 
Last edited:

CatPlayer

Commendable
May 12, 2016
41
0
1,530
0
You ought to be able to get 1366MHz @ 1100mV. The GloFo 14nm process freq/voltage curve has an inflection point at ~950mV. Anything below that is going to be a flatter curve than anything above.
Also, 925mV for 2000MHz VRAM should work (I can do 900mV).
The VRAM voltage is going to act as the lowest load voltage for any of the core P-states (it will still idle lower, ~800mV IIRC). ie, if you have VRAM @ 925MHz but core @ 1100MHz/900mV, you'll see that it only goes down to 1100MHz/925mV.

Turn "Frequency" to "Dynamic" and "Voltage" to "Manual"
If you post a screenshot or list the 7 core frequencies, I can generate a voltage curve to get you in the ballpark.
View: https://imgur.com/kA8sv61


Here you go brother, set to dynamic and manual, very useful information there, thanks. I will wait on your input now before changing any of these numbers.
 

tennis2

Notable
Nov 12, 2018
1,192
45
1,140
73
Try this on for size (obviously test for stability, each card is a little different).

600MHz/820mV
900MHz/870mV
1145MHz/925mV
1215MHz/945mV
1255MHz/970mV
1300MHz/1020mV
1355MHz/1100mV

VRAM = 2000MHz/925mV
Leave Memory Profile = "Auto"

I'd suggest checking the voltages are actually showing up like that on cold boot. I've had issues with machines not applying the voltages correctly on cold boot, but a single restart fixes the issue (and yes, you have to do this EVERY time you boot the PC). My personal machine does this.

Do you have a fixed refresh or FreeSync monitor? Using Frame Rate Target Control for fixed refresh and Chill for FreeSync is a good way to save additional power/heat/noise because it allows the GPU to downclock into those lower frequencies if you're playing a lightweight game where your GPU can crank out say 200fps but your monitor is only 60Hz. Likewise, Chill adds the ability to downclock the GPU in scenes/instances where there's little/no movement on screen.
IIRC, @ about 945mV, your GPU will probably only pull about 90W power draw (don't quote me on that, might be less, see what GPUz says)

I'm pretty certain the "Power Limit" slider does nothing once you're into manual voltages. I haven't changed it from 0% in a long time (I don't OC, I just undervolt while keeping the default clocks) and I can't remember the conclusion of my tests back when I investigated it. If you're on Auto voltage, the Power Limit slider will allow the GPU to suck more power and/or voltage. Therefore, TJHooker's advice about that is/was valid, but only for Auto voltage. Manual voltage, on the other hand, reduces the power usage, so you've got more headroom (at stock clocks).

From my experience, it seems AMD tested and found a fitting voltage curve for the majority of cards, then offset by about +50mV for the Auto voltages. Since the RX580 is just a further OC'd RX480, hitting power limits is a somewhat common occurrence. Coincidentally, the RX480 stock clock of 1266MHz sits pretty close to the 950mV voltage inflection point (kudos AMD!). You'll find that most chips have some sort of "safety factor" offset in this same vein (including CPUs, hence why you always see things like "I OC'd my CPU to xxxx @ stock voltage).
 
Last edited:

CatPlayer

Commendable
May 12, 2016
41
0
1,530
0
Furmark is dumb IMO, not representative of actual gaming load. Try something like one of the Unigine benchmarks (I like Valley) or 3DMark Firestrike/Time Spy, see what your clock speeds are like then.
I also did your suggestion and I had higher temperatures and lower clocks. I will install 3D Mark anyways, I will get back to you soon.
 
You get higher temps and lower clocks with Unigine than Furmark? Huh. Well, you can still just max out the power limit. Can also try reducing the voltage for P7 further, until it becomes unstable (will get artifacts or crashes in benchmarks).
 
Increasing power limit might result in higher temps, but it shouldn't result in lower clocks unless you also start overheating or something. It might just be Furmark, I think it ought to work better with a different benchmark (if the undervolt alone isn't enough to maintain a consistent clock speed).
 

CatPlayer

Commendable
May 12, 2016
41
0
1,530
0
Great news, with Time Spy it has maintained 1355 MHz after the undervolt and the very slight underclock, now what should I change to maintain Boost clock? For boost clock do I have to put it in manually? (it is 1386)

View: https://imgur.com/a/AOmE4Wx


Edit: added FireStrike benchmark to the album, this one struggled to keep clocked to 1355, it underclocked often to 1330s and 1340s while still having good temps.
 
Last edited:

CatPlayer

Commendable
May 12, 2016
41
0
1,530
0
By the way, I noticed this while playing resident evil 2 at 1440p, it would underclock to the 1100s and 1200s whenever the game got very demanding, it was not specifically furmark.
 

CatPlayer

Commendable
May 12, 2016
41
0
1,530
0
I increased the power limit by 5% and set the state 7 to 1386 MHz, something very similar to the 1355 MHz clock speed happened just with higher temperatures, where it would lower sometimes by 30-20 MHz but it still ddid the trick, regardless the 0.5% to 1.5% performance difference in this fluctuation is not much to make me worry about performance loss and I guess we found our sweetspot, regardless, I would like to achieve full stability at 1386 where it doesnt drop. Any tips?

View: https://imgur.com/a/AOmE4Wx


Here is the album with the new results.
 

tennis2

Notable
Nov 12, 2018
1,192
45
1,140
73
I agree that FurMark is a poor excuse for a stability test. I've actually had good results with Folding @ Home (believe it or not)

Seems odd that you've reduced voltage @ each perf state and it seemingly hasn't changed your situation. I'd recommend going for it and setting the power limit to the max (+50%) and see what that does.

I dislike the WattMan graph. I prefer to use GPUz.

In your RE2 testing, did you have FRTC/Chill off?
 

CatPlayer

Commendable
May 12, 2016
41
0
1,530
0
I agree that FurMark is a poor excuse for a stability test. I've actually had good results with Folding @ Home (believe it or not)

Seems odd that you've reduced voltage @ each perf state and it seemingly hasn't changed your situation. I'd recommend going for it and setting the power limit to the max (+50%) and see what that does.

I dislike the WattMan graph. I prefer to use GPUz.

In your RE2 testing, did you have FRTC/Chill off?
It changed the situation! Since I was having relatively estable numbers on 3D Mark, I only upped to 30% instead of 50% and I achieve this

http://gpuz.techpowerup.com/19/03/18/2v5.png

It was pretty stable, the clocks were good and no crashes whatsoever. Is there a way to make this more efficient to lower temperatures and thus, the fan noise? at 2600RPM it was not super loud but definitely could hear them very well which is something we should avoid, temperatures at 70% fan speed were ok though, as i believe 75ºC is good but I wonder if we can lower temperatures and thus general GPU noise. Any recommendarios? Thank you very much for the input.
 

tennis2

Notable
Nov 12, 2018
1,192
45
1,140
73
75C is likely the target temp set in WattMan.

Lower voltage = lower temps.
Lower clocks = lower voltage.

Also see my advice on FRTC and Chill from earlier.

205W seems high for 1355MHz/1100mV
 
Last edited:

Similar threads


ASK THE COMMUNITY