The problem lies at the contact-wire interface.
Look at the difference wire gauge makes: 10 amps with 16 AWG wiring drops to merely 6 amps with 22 AWG.
A dab of solder might be to the connector what thermal paste is for CPUs.

Did you mean:

                                      /------ pin in PCIe6 connector
12V from PSU ----- wire nut------- pin in PCIe6 connector
                                      \------ pin in PCIe6 connector

                                      /------ pin in PCIe6 connector
Ground from PSU -- wire nut------- pin in PCIe6 connector
                                      \------ pin in PCIe6 connector

(three PCIe 6-pin cables packed together at one wire nut)

Anyway, I don't see the advantage of choosing a wire nut over soldering... the latter pretty much eliminates the point of failure.
Both are 5 minute jobs...

Don't take my word for it, do your own math: the spec clearly says 1.5A per contact. Three contacts equal 4.5A per connector.
Moreover, the sata power connector was by a no means engineered for reliability. Instead, a lot of emphasis was put on ease of connection/disconnection.
You did notice this is a SATA - "molex" adapter, not a molex - PCIE adapter, right?

11.3V is already out of spec by 0.1V  
IDK how exactly you tested that PSU but do keep an eye on the voltage... it can be a sign that the unit is having a really tough time.
Let's cross the fingers and hope that the PSU can last a couple of days.

Not a DVD drive - the power utilization is nil when not actively reading a disc.

As it turns out, Art was correct - PCIE graphics manufacturers are prohibited from creating a galvanic link between PCIE power sockets.
Plug1 powers a set of voltage regulators and Plug2 - another set.
The VRM controller regulates those subsets independently and in general makes sure they behave.

If you feel comfortable with a soldering iron in your hand, rock on.
Twist the copper wires together and don't skimp on the solder, give them as much as they will soak.
Just don't forget to put the shrink-wrap insulation in place before soldering the wires together

PCIE 6-pin/8-pin plugs and contacts can be easily ordered online, you can make any kind of connector you need without that messy wire soldering.

Also, that 2*molex -> PCIE 6-pin you mentioned before looks up to the job.

EDIT::
Since you mention this is an old PSU, have you tested whether it will run with no load at 5V / 3.3V rails?
More importantly, won't it jump out of spec under the cross-load condition?
You might need to load the 5V rail with a power resistor.

Do you happen to have a multimeter lying around? Power up the card, and measure voltage on some other 12V cable connected to the same rail.
A double-rail design will likely have rail1 connected to the ATX12V connector (and perhaps the ATX 24-pin connector) and rail2 to all sata/peripheral connectors.
When in doubt and the user manual is inaccessible, open up the PSU and verify rail setup visually.
Measure again with the card having ran fully overclocked for a couple of hours. Better safe than sorry.

What a god-awful hack crimping two 20 AWG wires together at the molex⁽¹⁾ contact, where all the current will flow...
That sure is a fire waiting to happen.

And no matter what, don't ever use a sata-> molex adapter for powering anything more than a hard drive!
The SATA connector is rated much lower than that molex connector, at 4.5 amps max under ideal circumstances.



Notes:
(1) Actually, the four contact "Molex" connector should be called the 4 pin peripheral connector but everyone just calls it a molex.

Well done!

Not that I mean to kick you while you're down, A1, but I tried to raise the red flag early on only to have my posts removed.
You are an incompetent anyone with half a brain should keep away from.

You yourself admitted that a friend set up that pool for you.
You are the one who couldn't figure a trivial command-line miner out.

Isn't this your second pool that died a horrible death?
I sincerely hope there won't be a third, for the miners' sake.

Remember my previous posts about A1BITCOINPOOL Bulanula?
I tried to warn you...

The temps are as steady as ever, lately even a couple centigrades lower due to the winter conditions.
There are no clogged filters or malfunctioning fans. Fan speed is at expected 2800 RPM.

I might launch some older cgminer version using the old phatk kernel but honestly I don't think I care to.

Excerpt from my thermal log:

let's compare that with an entry from two months ago:

c0 as in Cayman0; c for clocks, t for temp, u for usage, f for fan. Exact RPM speed needed to be checked in cgminer.
This arrangement allows me to just grep the desired parameters from the weekly-rotated thermal log.
GPU parameters are logged every 6 minutes, every hour additional info such as CPU temperature, date and time is being gathered.

Roger that.
The startup peaks are insignificant - the real load begins once the OS boots up and the miners launch.

Precisely.

OP, there's nothing you can do about the GPU core quality. You can only try to drop the voltage manually, e.g. using cgminer, afterburner, or trixx.
Try to take the vddc downto 1.050V, then if it's still stable further down to 1.040, 1.030,...
Power consumption, and hence temperature, drops proportionally to voltage reduction squared - every milivolt helps.

By all means, you can try undervolting the second GPU too - you can't break it this way and may be able to drop the power usage by a nice margin.

Make sure to drop memory clocks as low as they'll go - you don't need fast memory for mining and the temperature drop can be significant when going from factory 1375 to 880 MHz.

Bummer.

Can you check the core voltages? The hotter card may use a lower quality GPU core requiring higher voltage to achieve the same clock as the other card.

If the voltages are exactly the same and the cards are the same make and model, the cooler may make poor contact with the GPU core as a result of manufacturing boo-boo.
Unfortunately, you'll likely void your warranty the moment you start messing with the cooler.

Who's the manufacturer of those two cards?

Without some fans, a case with the side panel removed would imply that the airflow is nowhere near fine.
You need far more than just convection to move the air around.

Hate to break your heart but cheap PC cases suck dicks when mining comes into play.
At the very least you should point two 120mm fans at those cards, e.g. one on the side panel (the more fannage the side panel can house the better) and one blowing from the front panel or from the bottom. Also, you need as much exhaust fan-power to make sure the hot air can be removed from the case double-quick.

I assume simply plugging the monitor into the weaker card didn't work.
Perhaps you can change card priority in BIOS by setting Init Display First to PEG2⁽¹⁾ instead of PEG1?

Notes:
(1) PEG as in PCI Express Graphics

Why? Are you trying to run some covert bitcoin mining operation on machines that don't quite belong to you?

Also, with five IO priorities introduced with Windows Vista any disk-heavy application, like a defragmenter or AV, can be made unobtrusive.

My ASUS 6950 DCII 1GB seems to be affected at last⁽¹⁾.

The card has been hashing since early September. Since late October the core speed was 942 MHz resulting in 389 MHash/s.
945 MHz would drop the card on its knees in a matter of hours. 942 MHz was stable for months... until this week.
The card hung a few days ago and twice yesterday.
935 MHz has been stable for a day now...

I was fully aware that setting the "cruise speed" so close to unstable clocks left me with no buffer zone - I did that more as an experiment into degradation progression than anything else.

I never liked this particular card much for its forced 125 MHz memdiff and locked core voltage resulting in power use of about 150W and temps in the high 70s.
Still, the degradation rate has been slower than I expected it to be.

Once the card is incapable of pulling 900 MHz I'll give it a second life in some gamer's rig provided no other issues emerge.

Just a data point for anyone interested in the subject.


Notes:
(1) Unless its just the latest cgminer pushing it harder than previous versions. The hash speed hasn't changed so I don't think the miner is responsible for the loss of stability.
     I don't feel like digging deeper into the matter or trying different kernels, I'll just drop clocks until stability is regained.

Is that really how misguided your opinion on Linux mining is?

It's the heap-of-fail AMD drivers are that made dummy plugs necessary. The problem was OS-agnostic though different driver versions solved it for Windows and Linux.

You just made me a BitMinter disciple, Doc:


That's with submit-stales enabled in cgminer.

Seriously, what do you need this information for and why Deepbit??
Plenty of 0% fee pools around.

Kindly allow me to elaborate on that:
Way back when we were all younger, life was easier: any PSU had discrete 12V, 5V, and 3.3V conversion circuitry.
A 300W PSU was considered a powerful, high-end unit. 5V was enough to feed a CPU. 5V was the primary rail of a PSU and had an evil twin: the -5V rail.
Any PSU could deliver up to X watts at the 12V rail, Y watts at 5V and Z watts at 3.3V rail, e.g. a 400 Watt unit being capable of 80W at 3.3V, 120W at 5V and 200W at 12V.
The load at any specific voltage rail was pre-determined by how powerful curcuitry had been installed.

Let's skip forward to 2012 - the dominant approach today is called DC-to-DC conversion.
For efficiency's sake the circuitry for transforming 220V⁽¹⁾ into lowly 3.3V or 5V has been ditched.
Instead those non-primary rails are being fed by the 12V circuitry and need only to regulate down from 12V.
Almost 100% (let's ignore the -12V and standby 5V rails) of the PSU's output is converted to 12V internally.
Success! - this gives you flexibilty previously unheard of: if you have a 650W DC-DC PSU, it doesn't matter whether your load pattern is 100W at 3.3, 130W at 5V and 400W at 12V or 630W⁽²⁾ at the 12V rail as long as you don't trip the internal overcurrent protection.

Mind you, that it's still the end-user's responsibilty to double check the topology and rating of any PSU before integrating it into such an edge case as a mining rig sporting 4 double-GPU cards.

The "50% load" rule-of-thumb advocates sound as if they spent a couple of years in a deep, dark dungeon... at least I hope they got some serious experience points out of it

Oh, and let me reiterate once more: disregard the peak rating⁽³⁾.

Notes:
(1) I'm not being Euro-centric. The 110V input voltage is being raised to 220V by the active PFC circuit or by the voltage doubler.
     The reason why PSUs achieve a couple per cent higher efficiency when plugged into 220V is that one circuit less needs to be used.
(2) I didn't use 650W as max load to accommodate for the discrete -12V and 5V standby rails often included in the max rating by those damned bean counters.
     Guys, if you're marketing a 650W unit, let me draw that 650W at the USEFUL voltages without overloading the PSU, ok? 650W should mean 650W, not 636.36251W
(3) Historically, the peak rating has been horribly abused to the dismay of many a user ending up with a dead PSU or even the whole machine.
     Huntkey PSUs were a blatant example: units with the overcurrent protection set at eg. 500W were being marketed as 500W units.
     Had the damned marketroids marketed them as 400W units which they clearly were many disasters could be averted.