XFX 6950 BIOS flash and voltages

[cicada]

I'm finally getting around to flashing my 1GB XFX 6950 ZDFC's (4 of 'em) to reduce the memory clocks and hopefully the temperatures along with it. 
(http://www.newegg.com/Product/Product.aspx?Item=N82E16814150552 is the card I've got)

These cards are a bit of a PITA due to a hard-locked BIOS that requires jumping pins 3 & 8 on the BIOS chipset in order to flash, but it is possible to do.  I've tried a couple of cards and was able to flash, though I was unable to unlock extra shaders on either, however this isn't my primary concern now.

Since there's no way to really reduce the memory clocks in Linux, I'm resorting to the BIOS flash to set default clocks.  At the same time, I figure it may be worthwhile to bump the voltages slightly for more stable overclocks; the best I can currently get is 885 GPU / 760 MEM before things go hairy.

My questions are these:  Does anyone know what voltage steps the XFX 6950s use?  If not, is there a simple way to figure this out other than trial-and-error?  This would be a pretty painful process of flash/test/flash/test if not. I'm willing to drop Win7 on a spare drive if there are good utilities to use to find this info out, though I'd rather avoid it.

I'm assuming the clocks and voltage stepping will be identical across all the various XFX 1GB 6950 cards, so if anyone has experience with best clocks/voltages on any of the other variants I'd be glad to hear it.  

[RyNinDaCleM]

Download Radeon Bios Editor! It will tell you exactly what the stock voltages are!

I believe 1.1v core, at least that is what my 2GB XFX HD695A-CNFC is at!

[cicada]

Do you mean min step supported by card (i.e. 1.0VDC, 1.025VDC = 0.025VDC step)?

If so I am not sure the min but I know 6850 (at least the 2GB version) supports voltage changes in 0.025VDC steps.

Yeah, that's what I'm looking for.  How did you determine this for the 6850?

[cicada]

I'm sorry I mean 6950 (6850 is typo). 

I found it by trial and error.  It likely will handle a smaller voltage step but 0.025 was good enough for my needs.

Thanks D&T.

It does appear the stock voltages for my XFX are 0.9, 1.0 and 1.1 for the three normal power states.

After some further research I've found that most of the other vendors are using 1.125 or 1.175 for the P3 state - it would seem that a 0.025 stepping is probably accurate. The MSI 6950s, which seem to be good OC'ers, use 1.175V so I'm going to try that voltage first.

I think that's well within tolerances for the GPU so it should be safe enough, and if it doesn't work out I'll get Windows installed and do some more testing as you've suggested.

[cicada]

Whelp at 830/350 1.175V it would boot fine and mine for about a minute before throwing the 'verify hardware!' error in poclbm - looks like it's not too happy there.

Win7 here I come

[RyNinDaCleM]

The 1.175v is the stock voltage for 6970's! If your VRMs are exposed (not cooled by the HSF), they probably got too hot. You'd need some little ramsinks to help cool them.

[cicada]

The 1.175v is the stock voltage for 6970's! If your VRMs are exposed (not cooled by the HSF), they probably got too hot. You'd need some little ramsinks to help cool them.

Well that'd explain it.  This is what I get for reading forum posts

I've got win7 running, any recommendations for utilities to change the voltages?  Afterburner doesn't seem to be able to do much on this card..

[jjiimm_64]

Since there's no way to really reduce the memory clocks in Linux,

I have 6 of these.  I run them at lower mems with the following command in linux

aticonfig --adapter=all --od-setclocks=880,440

I run them at 880 because they do not seem to be stable at the higher core clocks.

[cicada]

I have 6 of these.  I run them at lower mems with the following command in linux

aticonfig --adapter=all --od-setclocks=880,440

I run them at 880 because they do not seem to be stable at the higher core clocks.

Are you certain you're actually getting the 440 mem clock on XFX 6950 ZDFCs ?

aticonfig --odgc --adapter=all

should show you the current clocks and peak ranges.  I can set the memory clocks to whatever value I wish, but they run at 1300mhz regardless unless I also use AMDOverdrivectrl, at which point I can set them 125mhz less than the GPU clock.

If you are setting this successfully, and it's sticking to that clock speed when you fire up your miner, can you tell us more about your setup?  Specifically, which version of the catalyst driver you're using and whether or not you have to run AMDOverdrivectrl

[cicada]

Well it appears that 350mhz for the mem is just a bad value.

I was able to figure out the voltage control in Afterburner (it's an unlock option in the settings, silly me).

I've been running the card at 920 / 310, 1.125V, netting about 370MH/s in GUIminer for about half an hour without any issues.  I'm going to let this chug along overnight and probably most of the day tomorrow, and if it's still stable I'll be burning in these values.

At 350mhz for the memory, the card was very unhappy - artifacting and generally acting weird, and eventually locking entirely.

GPU speeds > 940mHz seem to lock the system instantly, but this could be partly due to the crappy no-name PSU I'm using in my test bench.

The temperature decrease isn't as extreme as I'd like to have seen, but this is probably due to the voltage bump.  GPU is showing 89C and VRMs are 93C - a bit warm but so is the room it's in, and there's no assisted airflow, just the card's fans at 90%.  This is basically the same temps I got with 885/760 @ stock voltage, but with a bump from 350MH/s to 370MH/s+

[RyNinDaCleM]

I thought I read somewhere here on these forums that the newer (11.6+) drivers allowed full under/overclockability in linux!?!

And, yes the "I am aware MSI does not support blah blah blah" EULA! I wrote a guide here for that! hehe

Well I'm glad you figured it out, and G/L mining!

[cicada]

I thought I read somewhere here on these forums that the newer (11.6+) drivers allowed full under/overclockability in linux!?!

Yes, the newer drivers do allow you to specify clock ranges outside of the 'current peak', however the 6XXX series still don't stick to the memory clock you specify.  Previously, aticonfig would tell you the values were 'out of range' if you specified clocks outside of the peaks.

I decided to bios-flash a slightly more conservative 885/310 and it's been running 7+ hours now without a hitch, so huzzah!

The PSU I'm using sucks pretty badly, and I'm only getting 11.19V on the 12V rail - would this increase temps on the card due to the higher amperage required?  I'm still at 82C right now with the memory underclocked.

[cicada]

Yes.  It also puts further load on the voltage regulators.  GPU and memory require tight regulation.  The looser the supplied voltage the more work you create for the voltage regulators.

I figured, just ordered a replacement PSU

This machine was running a 5830 that was sitting around 75C - moving that card temporarily to another rig with a nice Corsair PSU dropped it down to 60C!  I'm excited to get these 6950s running cool finally.

[CanaryInTheMine]

I had emailed PowerColor back and forth a few times regarding mem. underclocking on 68/9xx and they told me that AMD has put a restriction on how much the mem. clock can be lowered. "to preserve itegrity of operations blah blah blah"

I had another post with a bunch of info on flashing a 6870 x2 card.

Basically, it appears that if you flash the BIOS, the windows drivers will not work.  There are numerous reports that under linux it operates "fine" after a BIOS flash...