Consolidated Litecoin Mining Guide for 5xxx, 6xxx, and 7xxx GPUs

[Crazy rider89]

Hi guys, I'm trying to mining LTC using 2x5870 with this line:
cgminer -o stratum+tcp://coinotron.com:3334 -u desmofan89.58701 -p 1 --script

The command "--scrypt" aren't accepted by cgminer!

the output is: [2013-04-07 12:51:40] cgminer: --scrypt: unrecognized option

Then i tried with cgminer --enable-scrypt, but the output is:

cgminer: --enable-scrypt: unrecognized option

Even if the command "--enable-scrypt" is on the README file!

[1715670501]

1715670501

[1715670501]

1715670501

[1715670501]

1715670501

[1715670501]

1715670501

[Tejsei3]

Hi guys, I'm trying to mining LTC using 2x5870 with this line:
cgminer -o stratum+tcp://coinotron.com:3334 -u desmofan89.58701 -p 1 --script

The command "--scrypt" aren't accepted by cgminer!

the output is: [2013-04-07 12:51:40] cgminer: --scrypt: unrecognized option

Then i tried with cgminer --enable-scrypt, but the output is:

cgminer: --enable-scrypt: unrecognized option

Even if the command "--enable-scrypt" is on the README file!

--enable-scrypt is for when you are compiling cgminer from source. --scrypt should work for a binary compiled with the --enable-scrypt option. What is the source of your binary? Did you compile it yourself or did you download it? What version?

[rocks]

Hi, I am trying to switch my 5870 mining rig from BTC to LTC and am running into the problems below, if anyone has any ideas how to fix this I'd really appreciate it.

A 5870 has 1600 shaders, so my understanding is thread-concurrency should be set to around 1600 * 4 = 6400.

However if I set TC to 6400 with cgminer with,

Code:

cgminer --scrypt -o http://us-pool.give-me-ltc.com:8080 -u worker -p pw --thread-concurrency 6400 -I 13 -g 2 -w 256

I get the following error:

Code:

[2013-04-07 15:42:29] Failed to init GPU thread 1, disabling device 0
[2013-04-07 15:42:29] Maximum buffer memory device 1 supports says 134,217,728
[2013-04-07 15:42:29] Your scrypt settings come to 209,715,200
[2013-04-07 15:42:29] Error -61: clCreateBuffer (padbuffer8), decrease TC or increase LG

The only way to eliminate the error is to use TC of 1600

Code:

cgminer --scrypt -o http://us-pool.give-me-ltc.com:8080 -u worker -p pw --thread-concurrency 1600 -I 13 -g 2 -w 256

But with this I can only a) get around 250Kh/s per 5870 and b) get ~3 Hardware errors per minute (which I've read is caused by concurrency being too low).

My machine is Win XP x64 and it has plenty of ram.

Has anyone run into this before or have any ideas? I've had this system mining BTC very stable with no HW errors for over a year and switching back to BTC causes the HW errors to disappear so the rig is fine, its just the LTC settings.

Any suggestions would be very much appreciated. Thanks!

[tacotime]

execute command

Code:

setx GPU_MAX_ALLOC_PERCENT 100

close the terminal

open a new terminal and run cgminer

[rocks]

execute command

Code:

setx GPU_MAX_ALLOC_PERCENT 100

close the terminal

open a new terminal and run cgminer

tacotime, Thanks a lot, running setx did fix the error above.

For any others, after doing the above I then received the new error below.

Code:

[2013-04-07 19:15:15] Error -5: Enqueueing kernel onto command queue. (clEnqueueNDRangeKernel)
[2013-04-07 19:15:15] GPU 0 failure, disabling!
[2013-04-07 19:15:15] Thread 1 being disabled
[2013-04-07 19:15:15] Thread 1 being re-enabled
[2013-04-07 19:15:15] GPU 0 failure, disabling!
[2013-04-07 19:15:15] Thread 1 being disabled
[2013-04-07 19:15:15] Error -5: Enqueueing kernel onto command queue. (clEnqueueNDRangeKernel)
[2013-04-07 19:15:15] GPU 1 failure, disabling!
[2013-04-07 19:15:15] Thread 2 being disabled
[2013-04-07 19:15:16] Thread 3 being disabled

Changing "-g 2" to "-g 1" fixed that.

I'm now getting ~255Kh/s from each 5870, so guess it's time to try and optimize.

[Crazy rider89]

Hi guys, I'm trying to mining LTC using 2x5870 with this line:
cgminer -o stratum+tcp://coinotron.com:3334 -u desmofan89.58701 -p 1 --script

The command "--scrypt" aren't accepted by cgminer!

the output is: [2013-04-07 12:51:40] cgminer: --scrypt: unrecognized option

Then i tried with cgminer --enable-scrypt, but the output is:

cgminer: --enable-scrypt: unrecognized option

Even if the command "--enable-scrypt" is on the README file!

--enable-scrypt is for when you are compiling cgminer from source. --scrypt should work for a binary compiled with the --enable-scrypt option. What is the source of your binary? Did you compile it yourself or did you download it? What version?

Thanks, i solved (i have to add "--enable-scrypt" after the "./configure" command)

but now with this line:
cgminer -o stratum+tcp://coinotron.com:3334 -u xxxx -p x --script
i get a very low hash speed for 2x5870:
(5s):58.75K (avg):58.34Kh/s | A:2  R:0  HW:0  U:1.1/m  WU:144.2/m

It seems that cgminer are using CPU rather than GPU, How can i fix?

[FatMagic]

Thanks, i solved (i have to add "--enable-scrypt" after the "./configure" command)

but now with this line:
cgminer -o stratum+tcp://coinotron.com:3334 -u xxxx -p x --script
i get a very low hash speed for 2x5870:
(5s):58.75K (avg):58.34Kh/s | A:2  R:0  HW:0  U:1.1/m  WU:144.2/m

It seems that cgminer are using CPU rather than GPU, How can i fix?

Make sure you are using --scrypt and not --script

[chrcoe01]

Taco can you take a look at my problem quoted below?

Okay - need some help on this anomaly.

First rig - Sapphire 7970 standard edition with Vapor X, used settings suggested in Taco's thread - out of the box 650k/H per second.

Second rig - two (2x) Sapphire 7970 "GHZ Editions", used exact same settings in Taco's thread - out of the box - only about 350 k/H per second.

I have spent the last week in "newbie purgatory" so have been unable to discuss this issue.  Finally on the second rig, I was able to get 500k/H per second/card by putting aggression back to 13, and overclocking the memory to 1680Mhz.

Can anyone tell me WHY do two of the same brand/manufacturer, same model, have such huge variance in performance?

Why would the two "Ghz editions" VASTLY under-perform the standard one?

Does anyone have any suggestions experience with a Sapphire "GHZ Edition" that could explain the variance?  The standard one was plug and play @ 650k.   The better ones have taken a week of tweaking to even reach 500k/H, and NONE of the recommended settings worked.  I know it is not bad cards because they both perform the same - what is different about the "GHZ editions" that is hurting them so much, and made them so finicky?

First rig is Q6600 65nm, 4GB DDR2, PCI-E 2.0, Win7 Pro 64, Sapphire 7970 - 650kH/s

Second rig is 22nm Quad i5 Ivy Bridge, 8GB DDR3, PCI-E 3.0, Win7 Pro 64, 2 x Sapphire 7970 "GHZ Editions" - 990kH/s (both cards)

I followed all instructions in this thread.  The second rig would only do about 300 per card until I overclocked the RAM on both cards to 1700Mhz and dialed back aggression to 13.  Any variance from that setting and I get LESS than 500kH per card.

I would REALLY like to get the potential out of these supposedly "better" Sapphire 7970 Ghz editions.

this might be the problem I am having then ...

I have the vapor-x GHz edition because it was supposedly better and I am capped at 500kh (for each of my 4 cards) :/  I am running v12.10 drivers and 2.8 APP-SDK, my first thought is to uninstall 2.8 and install 2.7 because I read this was supposed to be better for mining.  But I had already finished setting everything up and just wanted to go to bed

if the standard cards are in fact quicker, then I am pretty pissed obviously since they would have saved me a few bucks each.  I might exchange them since they are recent purchases if we think the GHz edition is the root cause

as probably will be requested for everyone's use, the system is running linux and I have Sapphire Vapor-x GHz editions (4 of them).  I am waiting on my riser cables to get all 4 running at once, but I have tested each and currently only mining on two.  As stated, I was able to get them to max at 500kh with 975/1710 clocks, thread-concurrency 8192, I believe the voltages came stock at 1.25v .. I noticed the vanilla cards seemed to all have lower voltages, or is that my imagination?

Perhaps the higher voltage causes the mining speed to decrease (doesn't make sense, but neither does anything else that's happened with them).  Since I only am running two until my risers get here, maybe I will just exchange the other two for vanilla vapor-x cards.  If I can get those doing that they are supposed to, it would be worth the hassle for an extra ~150kh per card.

based on FullLife's post:

So after reading through this whole thead, I thought I'd contribute my setup and config.

Rig:
i3-3225
Gigabyte G1.Sniper 3 mobo (in hindsight, I should have bought a different board, this one is overkill)
8GB DDR3
(4) Gigabyte 7970s link (in hindsight, I should have bought either vanilla 7970s or 7950s, these are not worth the extra $$)
1200W Coolermaster Silent Pro Gold ps (I can't say enough about this power supply, it's excellent)
250GB Samsung 840 SSD
(2) WD hard drives (500GB, 2TB)
Windows 7 Ultimate SP1 64-bit

Setup:
So I had all 4 cards running in my case, but with them being all crammed together, it was way too hot.  So I pulled everything out and sat it on a rack.  I have 2 of the cards plugged into the board, a 3rd card plugged in via 16x riser cable, but couldn't get the 4th card to work with another 16x riser cable. So, I ordered a 1x to 16x riser cable and hopefully this will solve my problem.

In the meantime, I'm running 3 cards and getting pretty nice hashrates.  These are my settings and results using cgminer 2.11.3 and AMD Catalyst 13.1 drivers:

--intensity 13 -g 2  715/710 = ~1500Kh/s
--intensity 13 -g 2  925/1000 = ~1950Kh/s
--intensity 13 -g 2  1000/1206 = ~2100Kh/s
--intensity 13 -g 2  1022/1244 = ~2150Kh/s

Other settings:
--thread-concurrency 8192 -w 256

I've literally spent most of a entire week playing around with settings and trying to find the sweet spot and I feel like I'm pretty much there.  As you can see, I actually underclocked my cards to get these hashrates.  Stock clocks are 1100 core/1500 memory.  Unfortunately the default voltage is about 1.2 volts on my cards and its locked (@#&($$^(@&$!!), otherwise, I'd be able to save 100-150W easily by undervolting.



I'm seriously thinking about getting rid of these cards and getting 7950s or 7970s that don't have the voltage locked.  Having the voltage locked is such a huge drawback, I'm just wasting too much power right now.

I may try some of what he did because he has the gigabyte GHz edition cards and they seem to be hashing just fantastically.  I'm not sure if the voltage is locked or not on mine, but it might be worth checking this out when I get home to try it.  I can vouch for the fact that any intensity other than 13 on these GHz edition GPUs seems to be worthless.

sorry for a long post, I just wanted to get my thoughts documented on this in case it helps me (and others) later

edit: I forgot I set it up to connect remotely I set it to 1022/1244 (and played with other values around that) it seems that this was able to get me 20kh more, so I'm at 520k now... but I'm running drivers 12.10, maybe the key is 13.1 with these GHz cards, that will be my next thing to try, but I can't do it remotely.

[Crazy rider89]

Thanks, i solved (i have to add "--enable-scrypt" after the "./configure" command)

but now with this line:
cgminer -o stratum+tcp://coinotron.com:3334 -u xxxx -p x --script
i get a very low hash speed for 2x5870:
(5s):58.75K (avg):58.34Kh/s | A:2  R:0  HW:0  U:1.1/m  WU:144.2/m

It seems that cgminer are using CPU rather than GPU, How can i fix?

Make sure you are using --scrypt and not --script

Yes, it's "scrypt", i've only written wrong on the post.

[chrcoe01]

does it show the GPUs in the list once it starts running?

it should show all the possible mining targets that it is setup for, so if you had 2 gpus it should show those on the list (in my experience anyway).

I had an issue where I had to rerun the aticonfig command once I added more GPUs or it would not recognize them.  Mine does not allow me to mine on the CPU however, so I don't know any more than that currently

[northcape]

Perhaps the higher voltage causes the mining speed to decrease (doesn't make sense, but neither does anything else that's happened with them).  Since I only am running two until my risers get here, maybe I will just exchange the other two for vanilla vapor-x cards.  If I can get those doing that they are supposed to, it would be worth the hassle for an extra ~150kh per card.

My limited understanding is that the cards have a feature called Powertune Boost which dynamically manages clockspeeds and voltages for maximum performance.
That may be desirable in a gaming scenario where the rendering load changes from scene to scene, but not so desirable in a constant 100% load scenario.
 
See: http://www.pcper.com/reviews/Graphics-Cards/AMD-Radeon-HD-7950-3GB-PowerTune-Boost-Review

My Sapphire 7950s have a little blue buttons on them that change the BIOSes and apparently also enable/disable Boost. I had the buttons enabled during the first few hours of mining/testing (thinking boost = more performance? yes please!) and noticed that the clock speeds jumped from 925/1400 to 850/1250 all the time. Just like the picture of Metro in the review above. (which is mislabeled with FPS instead of Mhz on the y-axis but you get the idea)

Once I disabled the buttons and switched back to the regular non-boost BIOS, my clock speeds and voltages smoothed out and the hashing performance improved dramatically.

As I’ve mentioned, I suppose the feature makes sense for games where you have rendering performance valleys and spikes that allow Boost to make the most out of the available TDP envelope, but it doesn’t make much sense under constant load.

[qbits]

Just wanted to let you guys know that 5770 seems to be a nice card for LTC, regardless if it is a 512MB version or a 1GB version. "setx GPU... 100%" is a must, cgminer 2.11.3 and amd 13.1 drivers, the example below is Win7, but I'm sure it would run just as well in Linux.
Clock, Mem and V are stock (850Mhz,1200Mhz,1.08V), possibly could do better with some experimentation. Please use thread-concurrency of 3000-4096, single gpu thread (-g 1) and intensity 19.

But the best part is that these cards don't consume much power (about 100W - 110W max) and they cost about €50 ($65) pre owned. Many owners will part with them cheaply as they are not capable of playing the latest games. They also use only one 6pin pcie power input so most power supplies will be able to handle them. I power two of these on a 450W Corsair power supply (and intel core2duo 8200).

So: kh / s / $ is very good: 3.0 kh/s/$ also kh/J figure looks great for this card: 1.86 kh/J
Comparing 7950 (assuming it can do reliably 600 kh/s) figures are: (assuming local best prices and no VAT)
7950: €240 ($310), 600 kh/s, 1.94 kh/s/$, assuming 300W, then 2 kh/J or close.

also please observe that almost any decent 1000W PSU will be able to power 8 5770 cards which yield 1'500 kh/s, but will strugle with 3x 7950 as you may see from my previous posts regarding Corsair HX1000 and 58xx cards. Once you pass 1000W mark then PSU start to become really really expensive.

Code:

 cgminer version 2.11.3 - Started: [2013-04-08 18:43:01]
--------------------------------------------------------------------------------
 (5s):436.5K (avg):384.8Kh/s | Q:841  A:965  R:79  HW:0  E:115%  U:2.6/m
 ST: 2  SS: 1  DW: 1677  NB: 228  LW: 3597  GF: 2  RF: 1  WU: 354.7
 Connected to coinotron.com diff 128 with stratum as user qbit.five
 Block: 292ffe17faadc63d...  Diff:13.5M  Started: [00:59:33]  Best share: 145K
--------------------------------------------------------------------------------
 [P]ool management [G]PU management [S]ettings [D]isplay options [Q]uit
 GPU 0:  74.0C 2845RPM | 188.2K/187.3Kh/s | A:479 R:38 HW:0 U:1.27/m I:19
 GPU 1:  82.0C  63%    | 198.2K/197.5Kh/s | A:486 R:41 HW:0 U:1.29/m I:19
--------------------------------------------------------------------------------

 [2013-04-09 00:54:53] Stratum from pool 0 detected new block
 [2013-04-09 00:55:00] Accepted 3bd5009c Diff 128/128 GPU 1
 [2013-04-09 00:55:00] Accepted 2edfadf9 Diff 139/128 GPU 1
 [2013-04-09 00:55:01] Accepted a30126b0 Diff 3.17K/128 GPU 0
 [2013-04-09 00:55:36] Accepted 1c10c139 Diff 210/128 GPU 1
 [2013-04-09 00:55:36] Accepted 12620f85 Diff 293/128 GPU 0
 [2013-04-09 00:55:42] Accepted 6e5e90ff Diff 148/128 GPU 1
 [2013-04-09 00:55:59] Accepted ea5f5c5b Diff 147/128 GPU 0
 [2013-04-09 00:56:38] Accepted c8597551 Diff 180/128 GPU 0
 [2013-04-09 00:56:46] Accepted 7e9a82ed Diff 582/128 GPU 0
 [2013-04-09 00:57:07] Accepted 6a2c3c5b Diff 825/128 GPU 1
 [2013-04-09 00:57:15] Accepted 243f216a Diff 132/128 GPU 1
 [2013-04-09 00:58:16] Stratum from pool 0 detected new block
 [2013-04-09 00:58:21] Accepted b22210cd Diff 286/128 GPU 0
 [2013-04-09 00:58:33] Stratum from pool 0 detected new block
 [2013-04-09 00:58:35] Pool 0 stale share detected, submitting as user requested
 [2013-04-09 00:58:38] Rejected d8c90be4 Diff 229/128 GPU 0  (Job not found (=stale))
 [2013-04-09 00:58:44] Stratum from pool 0 detected new block
 [2013-04-09 00:58:46] Pool 0 stale share detected, submitting as user requested
 [2013-04-09 00:58:51] Rejected c455c0da Diff 161/128 GPU 0  (Job not found (=stale))
 [2013-04-09 00:58:54] Accepted a9ae0b1a Diff 128/128 GPU 0
 [2013-04-09 00:59:08] Accepted 5cef24c0 Diff 243/128 GPU 1
 [2013-04-09 00:59:20] Accepted 73d6aa36 Diff 1.25K/128 GPU 0
 [2013-04-09 00:59:33] Stratum from pool 0 detected new block
 [2013-04-09 00:59:35] Pool 0 stale share detected, submitting as user requested
 [2013-04-09 00:59:40] Rejected 7d97a34a Diff 4.86K/128 GPU 1  (Job not found (=stale))
 [2013-04-09 00:59:42] Accepted 0f3c85e6 Diff 202/128 GPU 1

[qbits]

My current best setup is 925/1250 at 0.962V for 560 kH/s at 170W or 3.29 kH/s/W, after many hours of systematically mapping gpu clock (

3.3 kh/J is great man!! Have you done any work with 58xx cards?
Cheers!

[trogdorjw73]

I have a few questions regarding LTC mining in general:

1) 7970 GHz edition: my card comes with factory clocks of 1050/1500; using various thread concurrency settings and intensity settings, I think the best I ever manage is a measly 450khash. I have also had issues with getting GUIminer-scrypt to work properly on this machine -- I think the best I got was maybe 300-350khash. If someone has some suggestions, I'm all ears!

2) I have a laptop with A10-4600M that also has an HD 7670M in it. I can mine with both the iGPU (APU) and the dGPU using GUIminer-scrypt, but when I try to run cgminer directly, I can't get it to work at all. I tried -d 0 and -d 1 and both options just ended with cgminer crapping out with an error. Also, mining speed on both GPUs is somewhat slower than I'd like -- it starts at around 70-80khash but after a bit of work settings in at closer to 50khash. Plus, when my primary mining pool goes down (which is happening more often than I'd like), I want to failover to a second pool. I'm not sure if I can do that with any of the other LTC miners, but cgminer seems like the best bet anyway. Since I can get cgminer working with GUIminer-scrypt, it has to be possible to get it running from a command prompt. Where can I find the exact commands being executed by cgminer via guiminer?

Edit: Found the console command for the second part. Interestingly, cgminer 2.11.4 doesn't appear to work on the laptop, but the version included with GUIminer-scrypt works. Failover now implemented... and damn you OzCoin, get your stuff together and stay up for a week without problems or I'm switching all my 4Mhash/s to another pool!

[rocks]

So the best performance I've been able to come up with for a rig with two 5870s is a mere ~290Kh/s per card with the following settings.

Code:

cgminer --scrypt -o stratum+tcp://stratum.give-me-ltc.com:3333 -u xx -p pp 
--thread-concurrency 6400 -I 17 -g 1 -w 256 --shaders 1600 --gpu-memclock 1200 --gpu-engine 850 --vectors 4

I've read through this guide and tried many many combinations of the above settings, with limited results. This rig is a Windows XP x64 box.

It seems people have been able to get above 400Kh/s per 5870, which is a significant gap from the impact from changing the settings above and what I've been able to achieve.

Any suggestions or thoughts?

Thanks!

[Tejsei3]

My current best setup is 925/1250 at 0.962V for 560 kH/s at 170W or 3.29 kH/s/W, after many hours of systematically mapping gpu clock (

3.3 kh/J is great man!! Have you done any work with 58xx cards?
Cheers!

Just the 7950 but if I do I'll share it here. I think I mostly got lucky just lowering the voltage at the stock gpu clock and memclock. I looked all over for more efficient combinations and nothing was as efficient as 925/1250/0.962V.

[Crazy rider89]

does it show the GPUs in the list once it starts running?

it should show all the possible mining targets that it is setup for, so if you had 2 gpus it should show those on the list (in my experience anyway).

I had an issue where I had to rerun the aticonfig command once I added more GPUs or it would not recognize them.  Mine does not allow me to mine on the CPU however, so I don't know any more than that currently

if i write: cgminer -n, i get:
 [2013-04-07 15:14:09] CL Platform 0 vendor: Advanced Micro Devices, Inc.                   
 [2013-04-07 15:14:09] CL Platform 0 name: AMD Accelerated Parallel Processing                   
 [2013-04-07 15:14:09] CL Platform 0 version: OpenCL 1.1 AMD-APP-SDK-v2.4 (595.10)                   
 [2013-04-07 15:14:09] Platform 0 devices: 2                   
 [2013-04-07 15:14:09]    0   Cypress                   
 [2013-04-07 15:14:09]    1   Cypress                   
 [2013-04-07 15:14:09] 2 GPU devices max detected   

I think that the gpus are correctly configured.
I will try to update the driver.

[northcape]

My current best setup is 925/1250 at 0.962V for 560 kH/s at 170W or 3.29 kH/s/W, after many hours of systematically mapping gpu clock (

3.3 kh/J is great man!! Have you done any work with 58xx cards?
Cheers!

Just the 7950 but if I do I'll share it here. I think I mostly got lucky just lowering the voltage at the stock gpu clock and memclock. I looked all over for more efficient combinations and nothing was as efficient as 925/1250/0.962V.

I have four Sapphire 7950s and can confirm that these are by far the best settings I’ve found so far, thank you for sharing them.
I run four cards at 925/1250/0.950, although that 0.950 is set via --vddc in cgminer, and GPU-Z shows ~0.996V for all GPUs. So  I guess they are actually not undervolted at all.

I could probably squeeze some more efficiency out of the system by setting each card to its minimum stable voltage, but mapping the correct cards between cgminer, GPU-Z and Trixx/Afterburner is too much of a hassle.

[chrcoe01]

does it show the GPUs in the list once it starts running?

it should show all the possible mining targets that it is setup for, so if you had 2 gpus it should show those on the list (in my experience anyway).

I had an issue where I had to rerun the aticonfig command once I added more GPUs or it would not recognize them.  Mine does not allow me to mine on the CPU however, so I don't know any more than that currently

if i write: cgminer -n, i get:
 [2013-04-07 15:14:09] CL Platform 0 vendor: Advanced Micro Devices, Inc.                    
 [2013-04-07 15:14:09] CL Platform 0 name: AMD Accelerated Parallel Processing                    
 [2013-04-07 15:14:09] CL Platform 0 version: OpenCL 1.1 AMD-APP-SDK-v2.4 (595.10)                    
 [2013-04-07 15:14:09] Platform 0 devices: 2                    
 [2013-04-07 15:14:09]    0   Cypress                    
 [2013-04-07 15:14:09]    1   Cypress                    
 [2013-04-07 15:14:09] 2 GPU devices max detected    

I think that the gpus are correctly configured.
I will try to update the driver.

I think i had to reinstall the driver after I booted the system with all cards plugged in.  The driver would not auto activate all cards.  So if you install the drivers with 1 card plugged in, then restart and plug all of them in, I had trouble getting it to show them all.  Then i reinstalled the drivers while the system had all cards plugged in and it worked after I re-ran:

sudo aticonfig --adapter=all -f --initiall

Perhaps the higher voltage causes the mining speed to decrease (doesn't make sense, but neither does anything else that's happened with them).  Since I only am running two until my risers get here, maybe I will just exchange the other two for vanilla vapor-x cards.  If I can get those doing that they are supposed to, it would be worth the hassle for an extra ~150kh per card.

My limited understanding is that the cards have a feature called Powertune Boost which dynamically manages clockspeeds and voltages for maximum performance.
That may be desirable in a gaming scenario where the rendering load changes from scene to scene, but not so desirable in a constant 100% load scenario.
 
See: http://www.pcper.com/reviews/Graphics-Cards/AMD-Radeon-HD-7950-3GB-PowerTune-Boost-Review

My Sapphire 7950s have a little blue buttons on them that change the BIOSes and apparently also enable/disable Boost. I had the buttons enabled during the first few hours of mining/testing (thinking boost = more performance? yes please!) and noticed that the clock speeds jumped from 925/1400 to 850/1250 all the time. Just like the picture of Metro in the review above. (which is mislabeled with FPS instead of Mhz on the y-axis but you get the idea)

Once I disabled the buttons and switched back to the regular non-boost BIOS, my clock speeds and voltages smoothed out and the hashing performance improved dramatically.

As I’ve mentioned, I suppose the feature makes sense for games where you have rendering performance valleys and spikes that allow Boost to make the most out of the available TDP envelope, but it doesn’t make much sense under constant load.

I think I had them disabled already, but just for kicks I enabled one of the buttons on one card .. it destroyed the hashrate and now cgminer reports it as "sick" whatever that means

as an update, before i tried to push that blue button, i upgraded to 13.10 drivers and it got me about 5 more kh so I was at 525kh for each card... now i'm wrestling to get all 4 working again ... don't push that button!

[mutikainen]

this was awesome, thanks!