Long runs (8-12 hours on fastest card)

Presently, the fastest single GPU card is the GTX680.
On XP the GTX680 takes the following times for different Long WU's:

SDOERR_2HDQd ~25,700sec (7.1h) 135,000
NOELIA_klebe ~27,500sec (7.6h) 127,000 (yes, there is a slight credit disparity, but you could argue that the SDOERR WU's are the ones that overpay)
NATHAN_KIDc22 ~36,500sec (10.1h) 167,550 (smack in the middle of the 8-10h estimate)

NATHAN_dhfr36 ~13,600sec (3.8h) 70,800 {these are the old NATHAN WU's and would normally have been in the short queue if there wasn't a shortage of WU's}
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Linux is the fastest OS, then XP. 11 of the top 20 machines use Linux or XP. So many of the top/elite crunchers appear to go out of their way to accommodate the project (and their credits). Perhaps this will continue.

The Linux/xp popularity drops when you expand to the top 250 systems, it's 65/250 - so 26%. However it remains at that even when you go past the 1000th top system - from 1000 to 1100 its 27% Linux and XP.

I have no doubt XP usage will continue to fall with time, but Linux seems to be picking up a lot of the slack (vaguely remember looking at this in the past).

The non project elite crunchers tend to use the less productive Operating systems and this continues through to the occasional cruncher and crunchers with bad setups or entry level cards.

While these lesser systems still belong to crunchers, how much work they do/can do, how important is it to accommodate them, and how much effort it takes is really a question that the project team has to continuously ask itself.

Ideally we could all crunch what we wanted, without problems, but this is more of a hands on project that is always changing. While I don't have the answers (just opinions)t I do have some data that's worth noting:

GPUGrid's total daily RAC is 191,926,972.
The top 20 systems contribute 20,507,084 (just over 10%)

The top 20 crunchers contribute 43,518,798 (23.9%)
The top 100 crunchers contribute 92,956,103 (48%)

The 1000th cruncher has a RAC of ~22,000 and by the time you get to 2000, the RAC is only ~1200. You could get a RAC of >22K by just running one long WU every 6th day.

There is two ways to look at this data, depending on human resources; we don't have a big enough team to accommodate everybody so it's best to concentrate on those that help the most. Our team is big enough to develop and expand to better accommodate more crunchers and new crunchers.

I doubt that the researchers will immediately go by the non-mainstream GK110 Titan when setting task steps against complexity (which equates to task duration on the top GPU), though it is a GeForce GPU. At present the number of crunchers attached to this project with a Titan is likely to be a single figure. It would make more sense to go by the GK104 GTX680 and then the GK104 GTX770 (which might be 10 or 20% percent faster), after there are plenty of crunchers who have the GPU.

____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Then again, maybe the SDOERR WUs are actually longer than the NOELIAs but run more efficiently. Looking at my GPUs the NOELIA WUs are running at 88-89% usage and the SDOERR WUs are running at 92-94% usage.

XP is 11% faster than Vista, W7 and probably W8 (though I haven't measured it).
Linux is faster than XP (up to 5% if it's setup correctly, but it might vary a lot on the system).
I've seen a GTX660Ti on Linux that is 25% faster than my FOC GTX660Ti on W7.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

skgiven, do you maybe have an idea if the performance difference is project dependent? So do other projects (and other gpu projects) also run that much faster on XP, in which case it would be an OS issue (I know that Win 7 is in general a bit slower than XP), or does this only happen with GPUgrid?

Beyond has covered the comparison to other projects.

The issue here is with the Windows Display Driver Model (WDDM) in Vista, W7, W8 and 2008+ servers. XP and Linux are not affected. It might be the case that there is even more of a gap now between XP and Linux when using CUDA 4.2.

From using MSI Afterburner it's obvious that Windows basically grabs a chunk of the display GPU's GDDR and keeps that for itself. It may control other GPU resources too. Don't know if this is a set/fixed amount, based on GPU model, CC or GDDR amount. It may be the case that if there are 2 GPU's in use, the one supporting the monitor under performs - but you would really need two identical GPU's and careful WU monitoring to determine that. Perhaps using the on-die GPU's prevents this 11% loss, but I can't test that (you need a motherboard that supports that to test it).

It was the Vista WDDM that introduced the idea of being able to restart the display driver, without restarting the OS.

http://msdn.microsoft.com/en-us/library/aa480220.aspx

As for relative performances, I have a reasonably FOC on my GTX660Ti. Sometimes it runs at 1202MHz, but at present it's operating at 1189MHz. If Linux is 5% faster than XP and XP is 11% faster than W7, then that's 16.55% faster, so to get to 25% faster you just need to up that by 7% (1267 or 1280MHz, without a GDDR increase). Readily achievable using liquid cooling, and possible with a GDDR tweak on air, with a good GPU. Another possibility is that on Linux the GPU utilization and power target is higher; 95% GPU usage over 88% is an 8% increase.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

I suspect it can be called a difference of "more closer to metal" with XP and Linux, whereas more abstraction in WDDM allows advanced features at the expense of performance.

the more complex the model the less efficient it becomes