Bitcoin Forum

Hello all

I just recently found out how dramatic an increase in efficiency undervolting can achieve. In my mining rig I'd like to optimize this, and preferably get it to consume around 250W in total. Right now it's consuming around 265W. I'm running with the following voltage/clock settings currently:

5870: 750 MHz / 1.000V / 97W / 343 MH/s / 3.54 MH/J
5770: 750 MHz / 1.010V / 63W / 168 MH/s / 2.67 MH/J

As you can the 5870 is significantly more efficient than the 5770. Then again it's also running at 1V vs. the 5770's 1.01V. But I simply can't get the 5770 to run properly at 1V at 750 MHz. At this voltage it keeps stalling in cgminer ("declared SICK") unless I run it at 700 MHz.

What are people's experiences with 5870's and 5770's and stable voltage/clock combinations of these?

By the way, the 5870 is an HD-587X-ZNFV V1.3 (http://xfxforce.com/thumb.ashx//4fecf3a2-1341-49db-a54a-c74f848589e0/RelatedGraphicCardSeries_RelatedGraphicCardModels1_RelatedGraphicCardProducts1/1024_x_768_jpeg_HD587XZNFX_1.jpeg,400,300) 5870, and the 5770 is this (http://eu.sapphiretech.com/mc/images/prods/343/11163-07_HD5770_512MBGDDR5_DP_HDMI_DVI_PCIE_C01_634418455064426122_200_200.jpg) Sapphire card, though I'm not sure if it has 512 or 1024MB RAM.

2 of my 3 5870s do 900 MHz @ 1.05V 
Both my 5850s will do 850 MHz @ 1V

^Cool. Good to know. I'm running at a lower voltage though. It seems even miniscule drops in voltage decrease power usage dramatically:

I just tried setting the 5770 voltage to 1V and downclocking it to 700 MHz. This seems stable. So these become the new figures:

5870: 750 MHz / 1.000V / 97W / 343 MH/s / 3.54 MH/J
5770: 700 MHz / 1.000V / 39W / 158 MH/s / 4.05 MH/J

What a dramatic increase in efficiency from a mere 0.01V drop!

Before reducing voltages, my mining rig would consume about 365W while producing 660 MH/s. Fiddling with the voltages, it now uses 240W while producing 500 MH/s. So that was 1.81 MH/J before, and 2.08 MH/J now. About a 15% increase in efficiency.

I'd love to get a Kill-a-watt device that supports retrieving wattage programatically in Linux, via an USB connection or something. That way I'd be able to make a script that can automatically change the clocks to achieve the best efficiency. It'd be interesting to see what the most efficient clock/voltage combination would be.

Can you try the 5870 @ 0.95V or 0.90V you likely will need to drop the clock some.

IM curious how you measure the power consumption of the cards alone.. ? Are you taking motherboard in to account?

Just tried upping the 5870 clock to 800 MHz. Here are the results:

5870: 800 MHz / 1.000V / 102W / 370 MH/s / 3.63 MH/J
5770: 700 MHz / 1.000V / 39W / 158 MH/s / 4.05 MH/J

So it definitely seems like the voltage is the most important factor in determining the power usage.

I'm trying out the 5870 at 600 MHz at 0.900V now. Will wait for the temps to stabilize and to see whether it's stable.

I'm simply measuring the wattage at the wall with either or both of the cards turned off in cgminer, and subtracting the wattages. So idle card power is not included in the card power consumption figures.

Yeah electrical theory tells us power changes at the square of voltage change.
So 90% of voltage = (0.9^2) = 81% of the power.

Theoretically if you could get a card to run on 50% of the voltage it would use about 25% of the power. :)

Power consumption is linear with clock rate.  So 10% reduction in clockrate should be ~10% change in power.

This is why overVOLTING is generally not worth it.  Lets say you overvolt 10% and get 15% higher clock rate.
Your power consumption will increase (1.1^2)*1.15 = 1.395  roughly 40% power increase for 15% higher performance.




If you want to be conservative the idle wattage of 5770 is <16W and the 5870 is <28W so subtracting that from rest of idle system would give you the "non" GPU idle load.

^ Interesting DAT. Thanks for the electrical theory lesson :).

The 5870 seems to be stable at 0.90V@600MHz. Here are the figures:

5870: 600 MHz / 0.900V / 61W / 273 MH/s / 4.48 MH/J
5770: 700 MHz / 1.000V / 39W / 158 MH/s / 4.05 MH/J

I'm running the 5870 at 650 MHz now to see if this is stable as well at 0.90V.

Would be interesting to see how high the efficiency can get.

Hmm. I just found out that it seems the leftmost hash rate next to the GPUs in cgminer might be an average-since-start. My previous figures might not be accurate then.

EDIT: The above hash rates are fine.

At 650 MHz the 5870 is about as efficient as at 600 MHz:

5870: 650 MHz / 0.900V / 66W / 295 MH/s / 4.50 MH/J
5770: 700 MHz / 1.000V / 39W / 158 MH/s / 4.05 MH/J

(Hash rate is now measured by a restarted cgminer)

How do you undervolt the cards? Trixx and Afterburner do not work for me (with the first the voltage is locked, and the second is not even able to read current voltage).

I just use cgminer (on linux).

You can try cgminer but not all cards have adjustable VRM.  Some cards (especially the cheaper ones) use a non-adjustable VRM to save money.

Does anyone know why my 5870 is stable at 800 MHz at 1V, while my 5770 is only stable at 700 MHz when running at 1V core voltage?
Is it just simple variance in the chips, and it might as well have been the opposite?

Using AMDOverdriveCtrl on Linux.

Manufacturing variance could easily explain that, but 5870 and 5770 also use completely different dies ("designs") and probably more optimized for yields than clocks.

How many different GPUs does AMD really make? I thought most of them were the same chip but they just disabled some of the stuff in the cheaper ones to make more money.

5970, 5870, 5850 and 5830 all use the same, relatively large "cypress" die (5970 uses 2 obviously).
57x0 (5750 and 5770 mainly) is a different, considerably smaller die called juniper. Below that you have few other derivatives still, if you want details:
http://en.wikipedia.org/wiki/Evergreen_%28GPU_family%29

From the 7970 thread, you take the difference between system mining wattage and the system idle wattage to get the card wattage.  It's not really exact, but it's at least consistent between different system configs.

Single 7970:
System Idle: 270 watts
System Mining: 388 watts
Mining: 925/150mhz,  865mv, 118 watts
550 MH/s (4.66 MH/watt)

Single 5970:
System Idle: 138 watts
System Mining: 268 watts
Card Mining: 680/166mhz, 950mv, 130 watts
622 MH/s (4.78 MH/watt)

System Mining: 250 watts
Card Mining: 605/166mhz, 950mv, 112 watts
550 MH/s (4.91 MH/watt)

System Mining: 235 watts
Card Mining: 605/166mhz, 900mv, 97 watts
550 MH/s (5.67 MH/watt)

Can anyone tell me how to get AMDOverdriveCtrl to set clocks and voltages on startup in Linux?
I've tried adding a script to be started in /etc/rc.local which looks like this:


But when I start up my mining rig and run aticonfig to see the clocks, they're the default ones, and cgminer reports that the GPUs are unstable.
Could it be that AMDOverdriveCtrl successfully sets the voltages, but then the clocks, for some reason are not set? the log file logs/overdrive.log shows that the script indeed has been run, but for some reason it either doesn't persist or some other program changes the clocks.

When I log in and run the above script manually, it sets the clocks and voltages fine, and aticonfig reports the clocks that I specified in my .ovdr files.

EDIT: Solved (https://bitcointalk.org/index.php?topic=59295.msg695394#msg695394)

Looks like the old 5xxx series still beats the new cards in $/mhash and mhash/watt.  Maybe the 7970 can be undervolted further.

Thats pretty damn impressive.  I mean BFL yet to ship FPGA only get 10MH/W.

I may undervolt my rigs this summer.  Keeping the garage cool with cards puking out 4KW of heat during August is tough.

I am thinking of undervolting now,  my power cables are already warm pulling too many amps thru them.

That is why I installed a 30A 240V circuit and 30A 240V PDU.  :)

The cable from the PDU to the L6-30R outlet is very thick.  With all the insulation I can't tell the exact size of the wire bug I have to guess it is 10 gauge or larger.

... yet the whole system only gets a measly 2.34Mh/J, why do I have the sneaking suspicion someones "idle" numbers are off?

imho,  idle numbers are useless.   measure watts at the wall, divide by Mhash.  eos

Who builds single graphics card rigs?

My idle numbers are fine.  AMD Phenom II 960T with all but 1 core disabled, and underclocked to 2.85GHz. It kind of jumps between 135-140W just sitting there doing nothing.  I can take a video of my Kill-A-Watt if you'd like?

I do :)  I sold all my h/w after we went under $3.00 per coin, but still kept my 5970 for BF3.  I still mine with it when I'm not busy pwning nubs.

Exactly, idle numbers are only used to establish a baseline.  "Zero out the scale" so to speak. 

DAT said, who builds single card rigs.?    not who mines with just a single card.

Yeah I guess nobody purposely builds single card rigs.  They're usually the result of just using whatever you have laying around until you make a decision to sink more money into it.  As the saying goes: A fool and his money are soon Bitcoin miner :)

I wasn't "bashing" you MousePotato.  Just pointing out that "low" system efficiency for a 1 GPU rig is silly.  The non-GPU load is ammortized over a single GPU.  Obviously it would be significantly better over 4 to 8 GPUs.

No worries, I understood what you were saying :)  Roadhog2k5 posted his results for the 7970 in both single-card and triple-card configuration but since I only have a solitary 5970, all I can offer is my single-card perspective for comparison.

But it doesn't actually zero it out, because you don't know the idle power consumption of the card. This means that adding another card will make complete system consume more power than the numbers in this thread indicate. The only thing these numbers can be used for is to decide if a PC with the tested configuration should be mining if it's already on anyway.

I thought about that, and I guess the only way to figure out how much power the card draws at idle is to have two identical cards.  Measure system idle with one card installed, then measure system idle with both cards installed.  The difference between the two is how much power each card consumes at idle.  I'm sure there's other factors involved that might affect the numbers, but that's about the simplest way I can think of.  Of course this doesn't apply to the 79XX series since it has the ability to hibernate GPUs that aren't currently active, so the MH/J for a multi-card 79XX setup are going to appear unusually low simply because system idle draw of a single 7970 is the same as that of a quad 7970 setup as I understand it.

You wouldn't need two identical cards if you can boot headless.  Boot and measure idle w/ no cards installed and with one card installed.

Also a logic puzzle would be to figure out the idle wattage of 2 different cards using no other cards and no headless boot (it can be done but involves multiple boots). :)




The hybernating 7970 uses much less idle power but it isn't 0.  I think AMD claim is <3W.  With 3x 7970 one could get the exact idle and hibernating wattage.

I ran the following to the mining shed (100+ feet):

12/3 plus ground
10/2 plus ground

at first I used only the black and reds as hots, which meant I had 3 hots, 2 neuts, 2 grounds.  I soon realized it was not enough:

so now
2 hot 10G off a 30amp double pole
3 hot 12G each off a 20 amp
2 neutrual returns (the old grounds)
gounds go straight to earth. (8foot copper)

this is all run from a subpanel that is fed with a huge gauge alum wire from a 100amp breaker.   I can tell you that every hot wire is warm to the touch, including the breakers!!

I am now looking into getting a separate meter :)

You are running 240V right to reduce amperage?

If you are then the major difference is likely the 100 foot run.  My outlet is ~5 feet from my distribution panel and the PDU has a 20ft power cord.

I'm still trying to figure out the answer to the logic puzzle.  :-[

Really.

System A = system w/ only card 1
System B = system w/ only card 2
System C = system w/ card 1 & 2

Boot system 3 times in the configurations above and record total wattage. (Wattage A, B, C).
Calculate the difference for
Wattage C - Wattage A
Wattage C - Wattage B

Wattage A = system idle + card1 idle
Wattage B = system idle + card2 idle
Wattage C = system idle + card1 idle + card2 idle

Thus
(Wattage C - Wattage A) = system idle + card1 idle + card2 idle - ( system idle + card1 idle )
(Wattage C - Wattage A) = system idle + card1 idle + card2 idle - system idle - card1 idle
(Wattage C - Wattage A) =  card2 idle


(Wattage B - Wattage A) = system idle + card1 idle + card2 idle - ( system idle + card2 idle )
(Wattage B - Wattage A) = system idle + card1 idle + card2 idle - system idle - card2 idle
(Wattage B - Wattage A) =  card1 idle

An example (and verification)

Unknown actual values
System Idle = 100W
Card 1 Idle = 15W
Card 2 Idle = 20W

Measured Results
Wattage A = 115W
Wattage B = 120W
Wattage C = 135W

Calculated Results
(Wattage C - Wattage A) = 135W - 115W = 20W  (graphics card 2)
(Wattage C - Wattage B) = 135W - 120W = 15W  (graphics card 1)

your 2 unknown card idle wattages are A and B, your system is X
measure: (X + A), (X + B), (X + A + B)
(X + A) + (X + B) - (X + A + B) = X
(X + A + B) - (X + A) = A
(X + A + B) - (X + B) = B

Ohhh yeah I'm an idiot.com/index.php?herp=derp.  I totally misunderstood the definition of "headless"

Anyone try this with 5830's yet?

+1

Unless the idle draw of the cards represents the same amount of logical work done, you can't just subtract them out to get the "mining" power consumed.
If one card idles hot doing tons of work that is not mining related, subtracting that from the final mining number misrepresents how much power that card actually draws in order to mine. Especially if it stops doing that work while mining. Conversely, if one card idles cold it won't have any idle wattage to subtract making its mining numbers seem like they draw more power.

The only numbers that can be compared are 100% mining vs 100% mining. Unless someone wants to map out exactly how much logical work each card does while idle and pro-rate their wattage for the extra work done...

Idle card does no work.  i.e. 0 mhash/s.

Look at it this way.

System at load:  300W
System at idle (including GPU idle wattage): 100W
GPU idle wattage: 10W

The reason we wan't to subtract the GPU idle wattage is to get the true GPU load wattage.

100W - 10W = 90W (system idle w/o no GPU).

300W - 90W = 210W (GPU full wattage at load).

Now we have apples to apples comparison, the GPU wattage at load.


We can also predict other system values.
system w/ 1 GPU  = 90W + 1*210W = 300W
system w/ 2 GPUs = 90W + 2*210W = 510W
system w/ 3 GPUs = 90W + 3*210W = 720W
system w/ 4 GPUs = 90W + 4*210W = 930W
system w/ 5 GPUs = 90W + 5*210W = 1140W
system w/ 6 GPUs = 90W + 6*210W = 1350W

Idle cards do no hashing related work. But it consumes watts. Therefore it does work. Explain exactly all of the instructions executed (like monitoring and answering driver polls, DMA channel, etc). If this work is accomplished only on an idle card and not on a mining card, then we need to account for that. Once those are accounted for we can compare to see if one card is providing say "full service with frills" to the OS, and the other card is "self service" while idle. That condition can skew the measurements. Especially with a card designed to go very cold while idle vs one that runs hot with idle cycles.

Apples to oranges.

Edit: A concrete example follows.

The AMD Phenom II X6 1100T (3.3GHz) consumes 20w while idle, and 109w under full load.
The Intel Core i7 2600K @ 4.4GHz consumes 5w while idle, and 111w under full load.

While under full load (in theory), the CPUs are not executing any idle cycles.
Also, voltage may be stepped down while idle and parts of the chip shut off further skewing the comparison.
Subtracting those 20 from 109  & 5 from 111 will not give you anything useful.

Hiho!

I have 2x 5830 and 4x 5770 cards in two rigs. Every card is overclocked and undervolted. My settings are:

"Sapphire HD5830"  core=900 memory=300 vddc=1.140
"XFX 5770"            core=880 memory=600 vddc=1.200
"Sapphire HD5770"  core=905 memory=300 vddc=0.960

"Sapphire HD5830"  core=910 memory=300 vddc=1.080
"ASUS HD5770"      core=925 memory=300 vddc=1.050
"Sapphire HD5770"  core=880 memory=300 vddc=1.010

Like you can see, there are many different setting on the 5770 cards. The XFX for example needs 600Mhz mem clock, no way to go more down. The Sapphire runs at 905 Mhz with only 0.960 vddc. I think you have to test a lot, till it's stable. Both RIGs need 650 Watts BUT with 3,5 HDD and the 100% Bug. So I think I can bring it down to 550 Watts @ 1,5 GHash. I'm looking forward to my new XFX 5970 Black Edition with 2x 920 Mhz @ 1.200 vddc ;-) (tested but not build into RIG yet).

Btw.: The great thing about my 5770s is, that all are "used" and I got them for around 60$ each on Sept. 2010! They are still a good catch I think.

Greetz
NetworkerZ

Even with 3x 5970s at 880 mhz (2400 mhash) around 900 watts its only  18 dollars a month to run them. The current payout is $343 a month. That leaves power at 5.2% of expense. Outside of keeping temps down it seems to make little sense to underclock. I realize power costs are higher in some places but it seems like if power was that high mining would be a bad idea to start with, maybe fpga would make more sense.

.9kw x 24 x 30 = 648.

$18 / 648kwh = $0.027/kwh  ???

You pay 3 cents a kilowatt-hour?

More reasonable cost are 10 cents a kilowatt-hour, for around $60 a month in electricity.  I think undervolted GPU mining will keep many systems up a while longer but my with FPGA's doing 400+ MH/s for 15 watts the day of the GPU is going to come to an end.

I pay .0202 cents a watt

You don't pay "per watt", you pay per watt-hour or to be precise, per kilowatt-hour (1000 watts per hour is $0.0202). Further, there are likely taxes and generation fees on top of that. That is a good rate however.

You pay $0.0202 per kilowatt-hour. At home I pay ~$0.32 per kilowatt-hour. My electricity costs at home are 16 times yours. At your power consumption rate that would put my expenses at ~80%. Undervolting can cut it from ~80% to ~40%, which makes it worth doing.

Damn, that's pretty good. I have one of those variable plans. It toggles from $0.05815/kwh during on-peak hours, down to $0.04273/kwh during off-peak hours. Of course I have to pay a flat "demand" charge of something like $10 per kW I use.  My power company likes to stack on little rinky dink charges that add up :(

What would be the theoretical correlation between stable voltage and clock pairs? I mean, if I decrease the voltage by 10%, should I also be able to decrease the clock by 10% and get a stable GPU? Or is it more complex, perhaps so much that trial and error is the only way to know? In any case, it'd be useful to have some rule of thumb to go by, even if it isn't completely accurate.

Currently, my mining rig is placed in a shed, where I would like the temperature to never go below 10C. Because of this, I'd like to be able to change voltage/clock dynamically, in order to control the power that my mining rig dissipates, thus acting as a sort of radiator with a thermostat. So it would be useful to be able to derive stable voltage/clock pairs from a known stable voltage/clock pair.

 Underclocking <300 even gives some performance!

 Here's what I've had before for my Sapphire Radeon 5770:
960/300/1.005v:
temp1: 67
temp2: 72
temp3: 70
fan: 50%
221.39

 And what now:
960/244/1.005v:
temp1: 66
temp2: 70
temp3: 69
fan: 48%
222.25 mhash/s!!!!

So,

I tried some undervolting on my cards, a couple had VRM chip, so I could change the voltage without any problems. Other cards didn't had any voltage control, so I tried to modify the BIOS of the card, to change the voltage of the card.

My test card was a Powercolor AX6770 with those default settings:
850 Mhz
1.2V

It worked partially, and I was able to modify the default Clock speed and the default voltage. I created a new BIOS with RBE, with the clock at 700 mhz and the voltage at 0.950. The BIOS has some "default" clocks, and I modified every one of them so they had 0.950V in their instructions.

When I booted the card on my mining rig, everything was in order. The card was by default at 700 Mhz, and cgminer showed me that the voltage was at 0.950. The thing is, while mining, the card didn't care about the voltage instructions in the BIOS. It was still consuming the same amount of Watts as before, with the stock BIOS. I could also clock the card at 900 Mhz without any problem, and with the same watts consumption as before.

Overall, the card just used the voltage it needed, ignoring the instructions in the BIOS.

Any of you tried something similar? I believe we have a deep topic on our hands, and it would be interesting to compile the informations of what you guys did.

Here's my results of my testing:
Diamond Radeon 5850
http://www.diamondmm.com/5850PE51G.php
Went from 2.68 MH/J to 3.51 MH/J
700 MHz at 0.88 V
260 MH/s

Sapphire Radeon 5830
http://devicegadget.com/hardware/sapphire-radeon-hd-5830-xtreme-review/3760/
From 2.07 MH/J to 3.16 MH/J
725 MHz at 0.95 V
215 MH/s

Sapphire Vapor 5770 (this thing is a beauty!!!)
http://www.sapphiretech.com/presentation/product/?psn=0001&pid=305&lid=1
From 2.13 MH/J to 3.42 MH/J
825 MHz at 0.95 V
178 MH/s

PowerColor 5870 (this one has an Arctic Cooler on it, because I broke physically the stock fan)
http://techiser.com/powercolor-radeon-hd-5870-ax5870-graphic-card-118806.html
From 2.33 MH/J to 3.51 MH/J
750 MHz at 0.95 V
330 MH/s

These cards have no VRM control:
Powercolor AX6770
http://www.shopping.com/power-color-powercolor-ax6770-1gbd5-h-radeon-hd-6770-1gb-128-bit-gddr5-pci-express-2-1-x16-hdcp-ready-video-card/info

Sapphire 6950
http://www.newegg.com/Product/Product.aspx?Item=N82E16814102914
Yeah, this one is surprising. You can use the Powertune to adjust the consumption, but it keeps the same MH/J.

Powercolor 5770
http://www.guru3d.com/article/powercolor-radeon-hd-5770-pcs-review/

I'm pretty sure all non-ancient graphics cards have voltage control. If they didn't, power management would be non-existent.

In Linux you should be able to use AMDOverdriveCtrl to control the voltage (within the limits set by the BIOS). I'm not sure which Windows programs achieve the same. Here's a list that provides some candidates: https://en.bitcoin.it/wiki/GPU_overclocking_tools

Many modern (mostly cheaper) models lack voltage control.  The VRM are not adjustable. Sure it wastes power but it makes the card $5 to $10 cheaper. 

Bingo! That's my problem. I can "change" the voltage, the software even indicates me the new voltage. But in reality, with a Kill-a-watt plugged in it, you see no difference. I can even input 0V, the software takes it, and the card run as normal.

I believe it could be possible to limit the voltage through the BIOS, but I'm don't know how yet.

Shows what I know. I thought adjustable VRMs were so cheap that no one would bother not putting one on the card. But then again I guess dynamically adjusting the voltage is a bit more complex than simply buying a capable VRM and slapping it on the card (like the additional software/BIOS development required).

Does anyone know if there are any 'cheap' 5970s out there that don't allow undervolting, and if so, which brands/models? I'm considering buying a 5970 but I don't want to risk getting one that doesn't undervolt (if any exist at all).

I have not seen any 5970, 5870s, 6990, or 6970 without adjustable VRMs.  I would imagine the same applies to 7970 & 7990s too.

It is generally the mid range and low end cards.  The reason for removing them has more to do w/ support calls & RMAs.  There isn't a lot of profit in midrange cards and most of that market is casual gamers.

I found one:
http://www.newegg.com/Product/Product.aspx?Item=N82E16814125316

I have a couple of those, and I can't change the voltage no matter what.

I have a couple 5970's (black edition) that I can't change the voltage on. I also have not been able to change the fans speeds either. Kind of strange, I have several of the cards and just two of them do this. Despite what pc there in. I had to put a fan controller on them and cut the blue fan wire to gain control of the fans speeds.

I have an xfx BE which I wish I never bought.  I would trade it for a reference ATI anyday.  Hell I would throw in $100.  Grr.  I don't have those problems but it is just "weird" when it comes to drivers.

Anyways.  You may want to check xfx website or google around.  xfx has a custom voltage mod tool maybe that will work.  I have a feeling xfx did "something" to the cards.  I am tempted to flash the bios to the "basic" XFX 5970 but am afraid I will kill it.  I will never buy a BE again.  AMD reference is good enough for me.

Isn't the XFX Black Edition 5970 a reference card? I know the XFX 5970 Black Edition Limited isn't a reference card, but the non-Limited version sure looks like all the others (except for some stickers).

Yeah physically it is but it seems to be an oddball when it comes to drivers and the like.  I don't know what they did to it but it doesn't like to play nice with other ATI reference cards.  I likely will ebay it and grab another ATI reference.  It has been (at least for me) a complete and utter pain in the ass.

How does this "weirdness" manifest, exactly? Would be nice to know since I'm currently looking for a 5970 to purchase.

Lots of things.  BAMT never sees the BE although it sees every 5970 just fine.  aticonfig -intial -f doesn't solve it.  If I add it to an existing windows machine it never sees the card.  It shows up in device manager but even catalyst control center (as well as GPU-Z, cgminer) never sees it.  The only thing that works is a clean install.  Installing drivers 12.1 causes BSOD but just with this card.  Remove the BE card and it installs just fine.  Insert BE and it BSOD on boot up.  Remove it and it works fine again.

It shows up in the "wrong order" in cgminer.  It makes the config file out of order relative to the display.   Every other card has the pair of GPU in the same order but the BE seems to put them randomly (it doesn't change but the order is random on each clean install).   Sometimes on a clean install it will show up as "disabled" in CCC.  It takes 2 reboots and it magically will show up again.  It doesn't like running at >835 Mhz even when watercooled w/ 50C core temps.  I can't change the voltage using cgminer (although I can w/ every other 5970 I own).  In cgminer it routinely stops showing temp or fanspeed (which is 0% because of watercooling).

Need I go on.  It is a piece of shit.  Maybe it was a lemon but weird when actually mining it mines just fine.  I have removed the waterblock from it and waiting on some TIM to put the stock heatsink back on.  If I can get it stable enough for gaming it likely is being ebayed. Honestly I have spent more time on this piece of shit then the other 14 graphics cards combined.  I should have given up a long time ago but I am stubborn.  Maybe i got a lemon but I am not taking any chances.

Would your XFX happen to have the rev. 1 BIOS? I've head that this BIOS from XFX is bad, and that - in general - XFX BIOSes are worse than other BIOSes. Take a look here:
http://forums.guru3d.com/showthread.php?t=329453

Unless you're done with trying to fix this card, putting a Sapphire BIOS on it might help. Here's a guide on how to do that:
http://forums.guru3d.com/showthread.php?t=329453

Without being a video card expert, I'm pretty sure that all the symptoms you describe *could* be caused by a poor BIOS.

Hmm.  Worth trying.  If it works I will give you a couple coins.  I already removed waterblock and my replacement thermal pads arrive today so I will remount the original heat sink, and try to flash it.  Honestly if I killed it w/ a bad flash I wouldn't be too sad.  :)

If it works reliably again well that is a nice score.  I will let you know.

You linked to the same thread twice.

^ Woops!

Here's the correct second link (about putting a Sapphire BIOS on the card):
http://forums.bit-tech.net/showthread.php?t=185918

Be careful though! Card-bricking is a possible outcome.

Damn it! I also have some 5770s which doesn't have programmable voltage controller.

I flashed BIOS with many different voltages, it showed new values nicely but the card just ignore BIOS and consume same amount of energy.

WTF >:(

Some lower end cards are not controllable.  The VRM regulator can't be altered by drivers, bios, or anything else (expect maybe some hardware mod).  The controller has a static value and the model chosen at construction time determines voltage. 

I know it is hindsight 20/20 and all that but generally this is why I stick w/ "upper half" cards (say 5850 or higher in 5000 series) and AMD reference designs.

My cards use a god damn uP6204BJ :(

I am looking for a way to hard-mod them. Maybe a reverse of pencil mod?

It's kinda strange, the MSI Hawk 5770 also uses that VRM chip. And MSI told that the Hawk series is very advanced and controllable

Dont blame me if it goes south:
http://www.overclock.net/t/997130/where-is-5750-powercolor-vtx-red-pcb-pencil-mod#post13226546

I did a pencil mod and thanks God, it did not go south!

After checking the datasheet, I decided to play with the voltage divider. The chip is programmable; but thanks to the laziness of Gigabyte engineers, they did not implement that feature.

Instead, a hard-wired voltage divider is used and that voltage divider will serve as reference voltage

http://desmond.imageshack.us/Himg28/scaled.php?server=28&filename=84097862.png&res=medium

I used a pencil and get the 5k Ohm resister went to 4.5k.

Voltage measured at the chokes changed from 1150 mV to 950 mV. Beyond my expectation :)