But are you using 2 GPUs on xinerama with one virtual screen only?

AMD OpenCL stupidly sees each "screen" as a device. That means that if you have 2 monitors connected to one GPU, you get two opencl devices. Presumably the converse is also true, seeing one device with 2 GPUs making one virtual "screen" with xinerama. I'm guessing if you force an extra output on the 2nd card, either by config or by adding a monitor/dummy plug to another output on it, you'll get two opencl "devices". Don't ask me who the moron in AMD was that thought opencl compute devices should be associated with screens and not physical devices. 

Best drivers for 5000 series were 11.6, but they won't even run on ubuntu 12.10 because of the newer Xorg.

The newer architectures based on GCN don't like vectors, so don't assume anything.

Well I don't know why you're taking this on yourself... I appreciate the gesture but it's not remotely trivial. Let me explain why it's now not that simple. As the avalon code was developed in secret outside of cgminer development, they worked on a codebase that is now redundant, hacking into it in a way where the only driver that works is for avalon. Cgminer's usb code and queueing has completely changed since then. To bring the avalon code into line would require large chunks of code to be rewritten to suit these changes. This is the danger of writing code out of the main tree. Now porting it to the new driver and queueing model and requires quite a bit of time effort and testing, and needs someone to support it (or not if it's abandoned). This is not in the scope of "lend me access to your mining hardware for a few hours over ssh".   I suspect Xiangfu will eventually be forced to keep his code in sync with the main cgminer git tree.

Yes I'm bitter about the whole Avalon experience, seeing them mine 10s of thousands of dollars worth of bitcoin each day with software mostly written by us, without them engaging us at all (except to get into a bunfight with Kano) or us earning a cent for our part, but it's not your fault.  

Alas you can't change the software on the unit, it takes uploaded firmware, so the only thing you can do is change the settings being passed to cgminer. While I'd be happy to tell your avalons to mine for me, I can't really improve cgminer to work better on them. Like I said, we weren't invited to the party.

As an aside, use the latest p2pool from git which should at least work with stratum with them and improve their reliability. In short, though, you lose on average .7 seconds of work per longpoll with Avalons. That's a fair percentage with average 10s longpoll equivalents.

Cause the cgminer devs aren't invited to the party 

cgminer already discards current work when the block changes. You have other problems if you're not getting accepted shares at all - look to your driver and sdk...

Try NOT ignoring the big fat warning at startup that says you need to install a new driver?

You know that's what happened with earlier versions... are you sure you're on the latest cgminer!?

No. Only Nvidia drivers are smart enough to do that.

I use the very first driver that supported the 7970, aka 11.12 with 7970 hotfix:
amd-driver-installer-8.921-x86.x86_64.run
I also use SDK 2.7 which seemed the best.

And as I said, I went by increments of 1 to the highest engine clock at that stability and intensity of 20 for a final:
Which gave 725kH without setting shaders, lookup gap, or thread concurrency.

Once again, your mileage may vary since all hardware is different, more system ram might help etc. Note also that setting the CPU frequency to performance instead of ondemand also helped a little.

It would be a good idea to implement the x-stratum header on your getwork pool while it's still up to redirect people to the stratum server. cgminer doesn't care when the header is added - as soon as it detects it it will switch to the stratum pool if it's alive.

And not enough power.

More of the same will happen.

Just let it go ffs 

Correct.

I still don't think we have an answer but more anecdotes to support the practice encouraged by my cgminer code based on my opinion. I had 4 7970s that ran at 100% load 24/7 for 12-14 months with variable fanspeed up to 85% to maintain temperatures constantly targetting 72-75 (as per the cgminer defaults when --auto-fan is enabled). The GPUs did not have a single problem and most have since been sold for a nice return.

No dotNet in the posix c code that is cgminer.

I'm guessing you're talking about windows. Windows does random things with the logging output it seems. Gotta love its (arse)soul.

win32 builds work on win64. However 2.11.3 is the latest stable release now and that's in the top cgminer directory.