Jump to content

[Mod] AMD 6970m voltage mod (*working*)


svl7

Recommended Posts

The AMD 6970m is an amazing card, great price performance ratio and really powerful. Only drawback I see is that there's no way to increase the voltage of the card through the VBIOS. This limits the maximum possible overclock and minimizes the benching capabilities of this card.

To solve this I decided to do some modifications directly at the hardware which allow me to increase and also monitor the core voltage of the GPU

How does it work?

The core voltage of the GPU gets controlled by a PWM regulator made by intersil, the ISL62883C.

It's a small 40 pin SMD chip, here you can see the pin configuration of it:

post-80-14494988554956_thumb.jpg

The exact purpose of each pin is described in the datasheet which you can find at the end of this post as attachement.

I won't go into details about how the chip exactly works, all that is needed for the voltage mod is Pin 12 (VSEN). This is the voltage sensing pin of the PWM controller, it monitors the current GPU voltage. As soon as the voltage drops below a certain level (e.g. because there's suddenly more load on the GPU), the chip will automatically give an additional PWM signal to bring the voltage back to the set level.

So if the voltage at Pin 12 is too low, the voltage gets raised. This means if you trick the chip into thinking the current core voltage is too low (even though it isn't), it will automatically raise it (and thus overvolt the GPU). This can be achieved by slightly dropping the voltage at VSEN.

The resistance between Pin 12 and ground is about 14Ω, putting a variable resistor between ground Pin 12 will allow lowering the resistance between ground an VSEN and thus lowering the voltage at VSEN. The resulting resistance is R1*R2/(R1+R2), R1=14Ω, R2 = potentiometer used for the mod.

A 500Ω potentiometer will do the job, I use a 1kΩ which is fine as well.

So all that's needed is:

- a potentiometer between VSEN and Ground

- a measuring point to monitor the GPU core voltage

post-80-14494988530235_thumb.jpg

Here you can see the conducting path that leads away from Pin 12 (red), possible solder spots for the mod are marked with green.

The parts are pretty tiny, you'll need a steady hand and a fine solder tip if you want to do this. It can be done without a magnifying glass, but it's good to have one for checking the solder joint afterwards.

On the left picture you can see the cable which leads to one end of the potentiometer soldered to the card. Could be better, but it works.

post-80-14494988553442_thumb.jpg

The other end of the potentiometer needs to be connected with ground. Since ground connection of the PWM chip isn't accessible (it's on the bottom of it) I had to look for a different possibility. I decided to go for an easy to solder spot, see the picture. This works fine, it has a big contact surface and soldering on a cable there is peanuts.

post-80-14494988530462_thumb.jpg

Such a voltmod requires a possibility to properly monitor the GPU core voltage. You can't just adjust it without knowing how much it gets changed (unless you want to burn your card). Since this isn't a desktop system it's not possible to simply do some measuring with a multimeter while the system is running, you need to solder on a cable to a measuring point for the core voltage.

The easiest possibility is to use one of the big electrolytic capacitors on the top of the card (the positive end obviously). Other possibilities are some of the ceramic capacitors on the other side of the die (which would probably be slightly more accurate). I tried this, but it's a PITA to solder there and I ended up going the easy way.

For ground reference you can simply use the ground cable which is connected with the variable resistor.

post-80-14494988530681_thumb.jpg

I also padded the VR with a high-end thermal pad. It'll run warmer than usual with the increased voltage. Hopefully it helps a bit.

On the other pictures you can see the cables leading from the GPU leaving the system. Black is ground, yellow leads to VSEN and orange is the reference for the GPU core voltage. The picture on the very right shows my voltage control center :D The potentiometer mounted in a breadboard, this allows to quickly remove the mod.

post-80-14494988553717_thumb.jpg

Here some quick tests. The first picture shows the system with the GPU idling, my multimeter reads 894mV and the VBIOS value is 0.9V (see GPU-Z)... perfect!

Middle: System at 100% GPU load (furmark), multimeter reading 1060mV, VBIOS value is 1.1V.

Right: GPU idle, I increased the voltage to 1V, card is still set to 0.9V (see GPU-Z)

So... yes, it works, but... there's a huge problem. When I did some very quick tests with furmark, the voltage gets automatically set back... I raised it further and further, until it didn't get set back anymore and I had a voltage of 1.2V... but the core clock dropped to 400 MHz.

I guess there's something limiting the maximum current, an overcurrent protection. Though what bothers me as well, was that according to GPU-Z, the VBIOS sets the voltage back to 1.0V when I raise it above a certain limit, so the VBIOS might be involved in this as well... I'm not sure.

More testing will come as soon as possible, I'll also try to force the clocks in the VBIOS. I really hope to get this working. If you have any ideas, please let me know. Also if something needs a more thorough explanation, just aks.

__________________________________________________-

EDIT: This definitely works!! Just did some 3dM11 tests at 915MHz and more. I though I solved the downclocking issue, but it's still there, randomly. It's probably caused by software issues (driver, ... ). Since I was really annoyed by this I decided to simply mod the VBIOS and set the 3d clocks and the battery clocks to 915MHz (and later more). This worked, the clocks got forced to this level, so this indicates that it isn't caused by the card, but by software issues.

Here's a 3dM11 test at 915MHz core, 1100MHz Memory:

post-80-14494991750441_thumb.jpg

You may realize that it's not really a highscore for the M15x. I had to leave the CPU at stock settings, otherwise I run into serious power issues. Take a look at the graphics score, that's more interesting.

As comparison, The Rev got 3965 GPU score on his best (submitted) 3dM11 test with the 6970m.

This one is at 935MHz, only the GPU tests:

post-80-14494991749723_thumb.jpg

I think there's even more headroom for the clocks, I upped the voltage to 1.25V, this should allow to get the clocks even slightly further, but I think I'll get limited by the VRAM at a certain point. Maybe I'll look for a voltmod for the VRAMs as well, but considering that there's already a power issue with this mod, it probably won't be of much use in the M15x.

Run out of time, sorry, I actually wanted to do some more testing, but the term started again today and I'll be pretty busy from now on. Also I had some other software issues, I guess my bench partition needs some work... reinstallation and fine tuning.

This was a very quick edit, just ask questions if something's unclear!

post-80-14494988529998_thumb.jpg

post-80-14494988553009_thumb.jpg

post-80-14494988553228_thumb.jpg

post-80-14494988553938_thumb.jpg

post-80-14494988554719_thumb.jpg

post-80-1449498855513_thumb.jpg

ISL62883C - 6970m.pdf

Edited by svl7
Messed up some pics... corrected
  • Thumbs Up 7
Link to comment
Share on other sites

Thank you, I just looked at some 6850 mods... which uses the same chip as the 6970m. Looks pretty similar when you look at the layout of the card. But I haven't found any information about OCP or similar limitations so far, the mod seems to work flawlessly on the desktop GPU. Though it uses a different voltage regulator...

Will do more research as time allows it, first of all I'll have to complete the first post.

Link to comment
Share on other sites

First post got updated with some explanations about the procedure, some more pics and description of the pics. Let me know if anything is incomprehensible, I was in rush, but I wanted to have this updated asap so I did my best in the short time I had.

Link to comment
Share on other sites

Did some more quick tests, this time with vantage. I don't have any auxiliary cooling methods here at the moment, and my room temps is at 28°C + ... the system heats up pretty quickly.

It seems that furmark also has some dips to 400/900 on standard clocks/voltage... though it happens more often with increased voltage. Could be caused by the temp. I did some quick vantage testing, not the whole test, only about the first 30 seconds of the first GPU test, it did run well at 1.2V, 855/1100, but suddenly the clocks and voltage dropped, then it came back for a couple of seconds and then the system shutdown.

Since it did run for the first 30 seconds I dare to hope that the issues are caused by heat. The overvolt/current protection would kick in within milliseconds, and it doesn't make sense, that there's no overcurrent/voltage during the first 30 seconds.

So I'll see that I can get some auxiliary cooling and do some more testing. It probably won't happen before the next weekend, but I'll see what I can do and I keep you posted.

Also there's some interesting stuff in the VR datasheet about the protective functions... I'll take a look at it, just in case.

I so want this to work :D

  • Thumbs Up 1
Link to comment
Share on other sites

stock volts i've had 845 1210 survive vantage and 3dmark2011. I wouldn't game at those levels though. SVL7 when i had a 5870M crossfire in an M17x R2 I was able to set up to 1.15 volts and reached 925 1270 at 1.15v. Could you try 1.15v over the stock 1.1? instead of the 1.2volts? Or no matter what the 1.15 will reset to 1.1? 1.15 should allow atleast 900 core and maybe 1200 stable?

Link to comment
Share on other sites

Yeah, I'll test it and see what I can do, but I need cooling, my room is 30°C, this lethal for an experiment like this. Testing is scheduled for Saturday... the wheather is actually too good for doing something like this, but I guess I can't wait. If everything goes according to plan I'll know a bit more by the end of the week.

I can try any voltage which is more or less sane, the potentiometer lets me adjust it pretty accurately. With the 5850m I was able to get 950MHz or similar (at 1.15V, stock was 1.0V) but the 6990m already starts at 1.1V, so it might need a bit more. I really hope it works, but as I said, without cooling I can't really do proper testing.

It didn't reset the 1.2V immediately, that's why I think it could be temperature related, the OCP or OVP should kick in almost immediately (1-2ms).

Link to comment
Share on other sites

Yeah, I'll test it and see what I can do, but I need cooling, my room is 30°C, this lethal for an experiment like this. Testing is scheduled for Saturday... the wheather is actually too good for doing something like this, but I guess I can't wait. If everything goes according to plan I'll know a bit more by the end of the week.

I can try any voltage which is more or less sane, the potentiometer lets me adjust it pretty accurately. With the 5850m I was able to get 950MHz or similar (at 1.15V, stock was 1.0V) but the 6990m already starts at 1.1V, so it might need a bit more. I really hope it works, but as I said, without cooling I can't really do proper testing.

It didn't reset the 1.2V immediately, that's why I think it could be temperature related, the OCP or OVP should kick in almost immediately (1-2ms).

Cool great work SVL7. My speeds on the 5870m's @ 1.15v were on stock paste so i'm sure what Stam has reached with his 5870M's on ICD 24 is a better figure on max clocks. So in theory 1.2v could be so much its constantly throttling. I'm hoping only half of that voltage would solve the thermal issue that may be holding this back. Yes the 1.1v is the default volts... and on the 5870m 1.1v for me brought me to similar ranges in core speeds as the 1.1v for the 6970m. One may require more Volts in a comparison of a same mhz overclock. But either way 1.15v brought the heat almost too hot as it was for my 5870m so i think you could pull this off at 1.15. there could be a limit like you said and anything over it would throttle so maybe if lucky 1.15v is under that limit. Either way this is a great mod with great potential.

Link to comment
Share on other sites

I really wanted / want to test this today, but my system is acting like crazy, I only get 400/900 on the GPU, even with the voltmod removed, powerplan is high, powerplay on, reinstalled Windows twice now and I've only seen 3d clocks once or twice. Everytime I've seen it running properly I prepared to do the testing, and then the clocks dropped back to 400/900, without having anything adjusted.

Once the clocks dropped when I started Trixx (but didn't change anything) and once when I unplugged the machine, but when I plugged the AC back the clocks stayed on battery mode, also once I've seen them dropping for no obvious reason.

This is so annoying, any help would be greatly appreciated, just any recommendations, no matter how stupid it is, I don't bother, I tried pretty much everything I could think of... my dry ice is melting away since hours and I haven't benched anything. :53_002:

Link to comment
Share on other sites

dude this sucks bro. had a similiar issue when i had the 6970. mostly after switching from igp to discrete, but for me a reboot always fixes it. you prolly already tried reflashing the card i'm guessing, see if you can change the 3d clocks to a slightly different numbers and reflash. i really hate seeing the dice going to waste.

Link to comment
Share on other sites

Haven't found a solution so far, even though I see 3d clocks once in a while, it doesn't seem to last long, even with the volt mod removed. No idea what's wrong... maybe it's the card, but this wouldn't make sense imo.

I haven't given up, but this cost me a whole day, might take a while until I can really solve this. Best would be to test it with another identical card, but that's almost impossible.

I tried reflashing the card, maybe I'll try a different VBIOS as well, but it's weird that it seemed to work a week ago and now it's completely acting up. Too bad I can only overclock the 3d profile... or is there a tool that allows to overclock the battery mode? (Like Nvidia Inspector for Nvidia cards... maybe ATI Tray Tools? Can't remember)

Link to comment
Share on other sites

  • 3 weeks later...

Finally had some time to test some more things, I reflashed the vbios and afterwards the clocks got back to 3d clocks as they should.... weird. Well, I can't complain, I hope it stays this way the next time I get some dice and then I'll see what I can do with the mod... keep your fingers crossed. I'm still eager to get this working.

Edit:

I just realized that I had the battery removed when the 3d clocks worked, so this might have been part of the issue... after all the clocks were on battery mode. This would rather indicate an issue with the mobo if this was the cause.

Can't double check it atm, I've just put a 5870m in my system which I bought from reborn2003, should be a beast of a 5870m, I'm eager to bench it, hahahaha. (And then I'll put the 6970m back).

Edited by svl7
  • Thumbs Up 1
Link to comment
Share on other sites

Yes... it works!! Just did some 3dM11 GPU tests at 935MHz, pretty cool, huh? :D And I think there would even be more headroom, I cranked up the voltage to 1.25V... yes, totally crazy - exactly how I like it.

Check out the first post for some more details, I did a quick edit and added some pics and information.

I wasn't able to do a vantage, it always terminated with an error just before the first test started, reinstallation didn't solve it

this, I think I have to redo my bench partition.

Also unfortunately I didn't have very much time for the testing, but at least I can confirm that it works. My previous limit was at about 860MHz core, now I run it with 935MHz and I bet there's still headroom, there's enough voltage :D

M17x r3 may have even better results, I had to leave out the CPU overclocking due to power restrictions... it's weird with the M15x, I can crank up the CPU like crazy (+300mV), but the GPU seems to be limited, actually the CPU seems to leech from the GPU voltage supply, but the GPU can't do the same vice versa... too bad.

Also I'm sure this will work for the 6990m as well, from the pictures I've seen so far it uses the same reference design as the 6970m, you only need to check the VR, especially the pins of it.

  • Thumbs Up 2
Link to comment
Share on other sites

Guest
This topic is now closed to further replies.
×
×
  • Create New...

Important Information

By using this site, you agree to our Terms of Use. We have placed cookies on your device to help make this website better. You can adjust your cookie settings, otherwise we'll assume you're okay to continue.