Reference 6950XT Underclock/Undervolt results

[ShadowChaser]

I know the 6950XT is pushed a little far beyond the optimal spot on the VF curve so for my daily use I simply dropped the PL by 10% and let it chug away happily at the 255W TGP that results. I got interested in seeing how much further down the VF curve I could go considering that a 10% drop in power is the maximum that AMD supports on this guy compared to the 40%+ on other GPUs.

 

I would be interested in seeing how other 6950XTs compare but I doubt people bought this card just to kneecap it with ridiculous limitations

 

I used Afterburner to make changes though Adrenaline would be fine as well, Time Spy to verify stability, and Furmark (1080p 8xMSAA) to get worst-case power draw. My testing was mostly limited by how much time the Time Spy stress test takes to run as well as how this card simply refuses to run all core loads at lower power (as is to be expected given the original PL limitation)

Undervolt	Target Clock	Power (W)	Core Clock	Core Voltage	Time Spy FPS Avg (20 pass)	%diff FPS	%diff Power	Furmark Power	Furmark FPS	%diff FPS	%diff Power
-10PL Stock	N/A	255	2235	1006mv	131	---	---	255	114	---	---
1080mv	2400	245	2330	956mv	137	4.58	-3.92	272	121	6.14	6.67
1090mv	2300	220	2235	900mv	132	0.76	-13.73	245	117	2.63	-3.92

 

The baseline is the -10% PL that I have been dailying and so the goal was to match performance while pulling less power. Should be easy, right?

My best result had the card running at 2235mhz @ 0.881v but it was borderline unstable so bumping it up to 0.9v cleared things right up. A power reduction of over 30W in a traditional rendering scenario is pretty solid IMHO and even under a torture test it still drew less power than the original power limit change.

 

This GPU really doesn't like to be worked hard under 0.9v I've found. Not sure why that is, perhaps it's so low that the vf curve just falls off. Maybe someone with more knowledge on how AMD's powerplay tables works could educate me on this as I've only really UC/UVed Nvidia GPUs in the past and they tend to be just as happy running at 0.85v as 1.05v.

[emosun]

1 minute ago, ShadowChaser said:

A power reduction of over 30W in a traditional rendering scenario is pretty solid IMHO

wow 30 whole watts huh?

[WereCat]

The issue is the limitation of the voltage curve and the fact the card will still try to reach the max voltage no matter how much you lower the slider so it will always max out the power. The only benefit is that the card will actually run lower voltage at lower clocks so on avg you will save some power but it also can cause instability in scenarios where the card is actually not doing much work while having no issues at max load.

 

(red is example of the voltage slider when you lower it in Adrenaline SW), I don't know for sure if it also lowers Vmin or not but I doubt it.

 

 

You can actually limit the max voltage and power with MPT and the card will never exceed those values so you'll be able to save a LOT more power with barely any performance degradation if you'll try to restrain it to stock clocks. And you can combine it with the Adrenaline voltage curve as well.

 

 

You can also save another 15-20W by just undervolting SOC. 

[podkall]

44 minutes ago, ShadowChaser said:

This GPU really doesn't like to be worked hard under 0.9v I've found. Not sure why that is, perhaps it's so low that the vf curve just falls off. Maybe someone with more knowledge on how AMD's powerplay tables works could educate me on this as I've only really UC/UVed Nvidia GPUs in the past and they tend to be just as happy running at 0.85v as 1.05v.

it's as simple as not being able to function properly due to lack of resources, same with car stopping because you used too little gas or let go of the clutch too hard/fast without having enough gas/motion

 

and since AMD isn't Nvidia, their power design is simply different, and so is efficiency, and other aspects like certain core counts, or if I compare a 6950 XT to a 3090 which are like 2% relative performance apart, 6950 XT runs at almost double frequency in memory clock, and clocks significantly higher at least in numbers, and all the other stuff like individual chip count, technology, generation, etc. (so much more)

[ShadowChaser]

1 hour ago, WereCat said:

The issue is the limitation of the voltage curve and the fact the card will still try to reach the max voltage no matter how much you lower the slider so it will always max out the power. The only benefit is that the card will actually run lower voltage at lower clocks so on avg you will save some power but it also can cause instability in scenarios where the card is actually not doing much work while having no issues at max load.

 

(red is example of the voltage slider when you lower it in Adrenaline SW), I don't know for sure if it also lowers Vmin or not but I doubt it.

 

You can actually limit the max voltage and power with MPT and the card will never exceed those values so you'll be able to save a LOT more power with barely any performance degradation if you'll try to restrain it to stock clocks. And you can combine it with the Adrenaline voltage curve as well.

 

You can also save another 15-20W by just undervolting SOC. 

I debated using MPT or not and it looks a little too complicated for a plug and chug sort of operation. Maybe in a few years when I retire this card and try and get as much out of it as I can. I'm not power or thermally restricted in my daily use case but going from 550W total system power down to about 400 without any perceivable performance loss is pretty good in my book. If I do something stupid like trying to stuff this gpu into a Fractal Terra again then maybe I'll invest the time and budget to do it right.

 

1 hour ago, emosun said:

wow 30 whole watts huh?

Shocking!

Jokes aside I dislike running this card at 350W+ but if you use that as a baseline...

[ShadowChaser]

34 minutes ago, podkall said:

it's as simple as not being able to function properly due to lack of resources, same with car stopping because you used too little gas or let go of the clutch too hard/fast without having enough gas/motion

 

and since AMD isn't Nvidia, their power design is simply different, and so is efficiency, and other aspects like certain core counts, or if I compare a 6950 XT to a 3090 which are like 2% relative performance apart, 6950 XT runs at almost double frequency in memory clock, and clocks significantly higher at least in numbers, and all the other stuff like individual chip count, technology, generation, etc. (so much more)

Yeah I just wish that you could edit the v/f curve in afterburner on amd like you can with nvidia. Guess I need to actually learn how to use MPT now

[podkall]

11 hours ago, ShadowChaser said:

Yeah I just wish that you could edit the v/f curve in afterburner on amd like you can with nvidia. Guess I need to actually learn how to use MPT now

you can, but it's like, managing to go 20km in a car for just 1 liter of fuel

 

and then trying in a different car going 20km for only 1 liter of fuel

 

these cars have different engines, and just because one can run certain mileage for X amount of fuel, doesn't mean the other car is capable of going similar mileage with similar amounts of fuel

 

• or that achieving same results will require same amount of effort

[Likwid]

I have rx6900 xt biostar and -80 mV, -5% power limit results in -10C while being a lot quiter and running 40w less

[Bagzie]

11 hours ago, Likwid said:

I have rx6900 xt biostar and -80 mV, -5% power limit results in -10C while being a lot quiter and running 40w less

Why would a 5% power limit drop 40w off the power usage?

[Likwid]

19 hours ago, Bagzie said:

Why would a 5% power limit drop 40w off the power usage?

-80mV mostly did the trick imagine if core runs at

 

1.2V at 250Amps you get 300w.

1.12V at 250Amps you get 280w.

 

Temperature drop also helps to reduce power.

 

Lower temp, lower power leakage.

 

TechYesCity has a great video about underclocking