Here's my opinion:
[humor]
I think everyone should recycle all their cards and immediately go purchase NEW, high-end, NVIDIA cards. Why? Because I own a small equity position at NVIDIA Corporation and your actions will cause the stock to rise.
[/humor]

Crunch on ...

Perhaps we should try to assemble a quick FAQ for new GPU crunchers?
Here's what I think I know, please correct me if I'm wrong.

Overheating may be a more frequent problem with GPUs than with CPUs because
* mid-range and high-end GPUs draw a lot more power than a CPU, and more power equals more heat. (High-end CPUs may draw 90-130 W, while high-end GPUs are 250-350 W.)
* GPU and case cooling is frequently designed on the assumption that the GPU won't see high load over extended periods. This is true for normal desktop use, but not when you run a graphics card for 24/7 number crunching. (There are stories of people having their graphics card catch fire running Furmark. Furmark is a special case, nothing in the real world will load the card that much, but it illustrates the problem.)

What do do:
- Install a GPU temperature monitor. For Windows, MSI Afterburner is nice (and works even if you don't have a MSI card), there's also SpeedFan and a whole host of others. Linux has a few command line alternatives, e.g. nvidia-settings and sensors.
- Ensure sufficient case ventilation.
- Get a card with decent cooling. Check reviews, there will be a section where they test card temperatures under load.

This is not so much a concern with low-end graphics, since they don't generate that much heat in the first place. But for high-end graphics, and for laptops with discrete GPUs, I would install a temperature monitor.

This is a deliberate Windows policy: Services are not allowed to access the GPU.
AFAIK there is no way to override this. So you can either do GPU crunching or run Boinc as a service, but not both.

AFAIK SLI and Crossfire is irrelevant to Boinc, it will treat each card separately in any case.
In short, it doesn't matter.

From what I've seen on various forums, mixed cards are nothing but trouble. I would stay clear if I wanted hassle-free crunching. (Merely having mixed cards in the box shouldn't be an issue, but getting crunching to work on all of them simultaneously can be tricky.)

As above, except that cards in the same family / with the same architecture should be fine. I think issues tend to arise when the cards need different drivers.
As an example, I expect that I can mix Radeon 6950 and 6970 without issues, but wouldn't hold my breath for mixing 6950 and 7950.

Noise
GPU coolers can be rather noisy under load. On laptops the options are limited to non-existent, but on desktops there are a few alternatives if you care about noise:
- Check reviews before you buy, many reviews include noise tests.
- Get an aftermarket cooler. ( Example)
- A "ghetto mod", stripping off the casing and strapping on larger and quieter fans, may reduce the noise to almost zero. (Caveats: Probably voids the warranty, track the temps closely afterwards to ensure that it works as intended. Example)

High intensity ATI/AMD GPU projects such as MilkyWay@Home, Moo! Wrapper and Donate@Home may actually draw slightly more current on the core VRMs (VDDC) than Furmark. Furmark would possibly cause greater total rise in card heat because it stresses the memory as well, but those projects can still quickly burn out a card's VRMs if VRM temperatures are not carefully monitored.

When the VRMs get too hot, it is necessary to improve cooling and/or reduce application intensity with <cmdline> parameters. If this is not done and cards are operated with very high VRM temperatures then card life will be reduced and in the worst cases it can cause card failure in a very short time.

A few years ago I quickly discovered the cooling of my old case is insufficient to cope with GPU processing so I run with both case sides removed. On hot summer days I aim a 30cm floor fan at the GPU.

VRM temperatures can be too high even when GPU core temperature is within an acceptable range. Some cards are more prone to VRM overheating due to poor component placement/cooling design. The HD 5970 is one of those.

Some more recent GPU models have protection circuitry that detects excessive heat/current in a number of components and throttles the card. Earlier models may throttle the card only on GPU core temperature and high VRM temperature does not trigger the protection circuitry.

The current draw and heat produced in VRMs is a function of how hard the GPU is working not just GPU load percentage. For example two different applications on a Cypress core can have the same 99-100% GPU load and one application can draw 38W on VDDC and the other application draw 75W. The application with the 38W current draw would have no problem with VRM temps as long as the GPU core temp was within an acceptable range. However the application with 75W current draw could easily overheat the VRMs, even when GPU core temp was within acceptable range.

This is why keen ATI/AMD GPU crunchers usually monitor VRM temps as well as GPU core temps.

It is possible the VDDC current draw on a WCG HCC GPU application will not be very high and so these precautions may not be so important here and monitoring GPU core temp alone will be sufficient for most except perhaps for those with heavily overclocked GPUs.

Good link coleslaw, and agreed on the GPU Forum...
Nice VRM/VDDC writeup kashie. Good to hear someone go beyond the GPU temps.

The HD 7750 looks like a good option for people with lesser PSU's; it won't need any extra power cables as it's only 75W (max). For crunching I would expect the quoted 55W to be correct for most projects and I doubt if any would go over 65W. In terms of value for money it also looks reasonable, but don't expect the performance of these cards to be immediately met by project apps; they may need to be honed.

If your PSU has a free 6pin power connector you might want the 7770. In terms of performance per monitory value and performance per Watt, this should slightly outperform the 7750 (by around 8% and 30% respectively). So it's worth the extra if your system can support it.

It's worth remembering that an HD 7970 has the same number of stream processors as two 7770's and a 7750 put together, but has a TDP of only 200W, so it's more efficient again. The problem is the expense and supporting a 200W GPU. Fortunately next month we will see two cards to fill the gap between the HD 7770 and the HD 7950; the 7870 and 7850... If anyone is considering two HD 7770's I suggest you wait a month - the 7850 (~$249) should be slightly more than twice as powerful as the 7770 but use slightly less than twice the power. A better buy in the long run. Partners might even be unleashed.

PS. Avoid the OEM 7000 series GPU's (7350 to 7670); these use 40nm Lithography.

Very interesting post there Maybe I want get one of the 7970 as they get a BIT cheaper. But I need to find a solution on the idea of having such a build run 24/7 or maybe have a standard CPU-build off the project. But in the end I think, that the efficency of the GPU-Card will make up for this big time. Prices may drop as Nvidia will come up with its new gen. Let's see if we have some GPU Apps here by then.