Exactly when did you do these tests?
How did you come to the conclusion that Win7 is slower for GPUGrid?
What assumptions did you make and what was the determination methodology?

Assuming your tests were solid and the results firm, I would speculate that it could be the case that it is slower on your particular system only; perhaps due to your AMD Athlon II X4 620 Processor not having L3 cache on that or perhaps it does not quite have enough RAM (64bit systems utilise more RAM, especially W7). You could well have unwanted services running (indexing services, defrags, aero, updates downloading, AV, you left a CD in, or it could be going into a low power mode). Perhaps you were even just using the computer more when W7 was running, or indeed perhaps your right, it is slower!

You should note that not all tasks are the same, some will take longer to run. Just look at your own results for your XP with the GTX 260; some tasks took 22000s and some 39000s. So if you ran 6 tasks on XP and 6 on W7 there is a fair chance that your average would not be an accurate conclusion. To be reasonably sure of your conclusions you would really need to run over 20 tasks on each system, from the same batch and see a constant and considerable change.
How much RAM do you have and what else is running?

This would lead me to believe you did your tests some time ago,

1697533 1056661 2 Jan 2010 5:15:54 UTC 2 Jan 2010 5:16:20 UTC Aborted by user 0.00 0.00 0.00 --- Full-atom molecular dynamics v6.71 (cuda23) 1692439 1056889 31 Dec 2009 16:20:59 UTC 31 Dec 2009 18:49:49 UTC Aborted by user 0.00 0.00 0.00 --- Full-atom molecular dynamics v6.71 (cuda23) 1690691 1056954 31 Dec 2009 4:11:44 UTC 31 Dec 2009 4:12:19 UTC Aborted by user 0.00 0.00 0.00 --- Full-atom molecular dynamics v6.71 (cuda23) 1690528 1057073 31 Dec 2009 3:03:22 UTC 31 Dec 2009 4:11:44 UTC Aborted by user 0.00 0.00 0.00 --- Full-atom molecular dynamics v6.71 (cuda23) 1687348 1056841 30 Dec 2009 7:05:13 UTC 30 Dec 2009 7:07:18 UTC Error while computing 2.10 0.06 0.00 --- Full-atom molecular dynamics v6.71 (cuda23) 1686462 1056695 30 Dec 2009 2:12:03 UTC 30 Dec 2009 7:02:13 UTC Error while computing 1.72 0.09 0.00 --- Full-atom molecular dynamics v6.71 (cuda23) 1685947 1057022 30 Dec 2009 0:35:14 UTC 30 Dec 2009 2:12:03 UTC Aborted by user 0.00 0.00 0.00 --- Full-atom molecular dynamics v6.71 (cuda23) 1229528 762768 9 Sep 2009 14:57:56 UTC 9 Sep 2009 18:04:22 UTC Aborted by user 0.00 0.00 0.00 --- 1162029 720687 26 Aug 2009 7:51:30 UTC 26 Aug 2009 16:23:18 UTC Error while computing 0.00 6.95 0.03 --- 936886 602381 8 Jul 2009 12:15:05 UTC 9 Jul 2009 14:49:12 UTC Aborted by user 0.00 62.94 3,977.21

You asked for an opinion and used vague wording such as, A While Back, Some Of Us Noticed, After Running Many Work Units...
You could have said, 4 months ago, many people noticed a 12% drop in points, added a thread link, said your machines had high end CPUs, plenty of RAM, large HDDs, and you ran 25tasks on each setup, stated the applications tested – or similar.
You could have also added that you work in IT, and then I would have assumed you did not make any obvious mistakes. That said, when W7 first came out, none of us were W7 experts, it takes time to become familiar with a new Operating System. Even after using Vista, I found that quite a bit had changed, and W7 is much more like Vista than XP.
Your tests may well be Valid but you did not supply enough detail for me to determine this and decide for myself. It takes facts, figures and confirmation to gain credibility.
When W7 first came out, and windows updates were allowed to run, it would automatically download an older NVidia driver and replace any newer one you installed yourself! This is because Microsoft only releases more heavily tested versions of other vendor’s software, so they are usually several release versions behind. This may have been the issue, but I can only suggest possible causes. You probably already know that XP is totally different in build than Vista and W7 which share similar kernals. This may also be the reason, as could a chipset driver, Bios version, or an automatic system configuration setting.
I think after SP1 is released would be a good time to review.

Have you tried this?

Advanced system settings, Performance, Advanced, change from Programs to Background Services.
I think I remember a card ran faster in a Win Server 2008 than on Vista. This may be why (the server even had a lesser CPU).

Might be worth a run overnight to test against previous task times.
Best to restart system after such changes.

Yeah, AERO has to be turned off to have any hope of similar performances with Vista/W7.
I dont think its automatically installed on 2008servers.

The Windows Desktop Manager is a bit of a hog, but you cant turn that off!
Lots of page faults, and its high priority. Lots of high priority tasks on Vista/W7, not as many on XP (and SP2 is faster than SP3)!

Perhaps a minimum installation of W7 would help?

Well, lets look at the difference between GPUGrud tasks and other tasks.

Firstly, GPUGrid uses NVidia cards (obviously), most other projects either use ATI or both (but are not so hot when it comes to NVidia cards)!

Most other projects have smaller tasks,

and most other projects use less CPU.

As the CPU usage on Vista/W7 for GPUGrid is set to below normal priority, while the many OS tasks are set to High Priority, and there are many, so they do keep interupting, perhaps that is the problem; not that the GPUGrid tasks are low, just that there are so many (and usually unnessasary) high Vista/W7 OS tasks that GPUGrid tasks dont get enough of a look in.

Well, the lower GPU utilization is a consequenze of the problem, not the reason.

Vista / 7 require completely rewritten graphics drivers. It could well be something small nVidia changed and possibly didn't notice, because it doesn't affect most apps.

When I switched from XP 32 to Win 7 64 I had a problem with MW on my ATI: performance dropped by ~30%. First I solved it by leaving one core free. That brought the speed back up to XP level and GPU utilization back to 99%. Now I've set "b-1" in Gipsels app_info to modify the polling behaviour. By doing this the app waits the estimated time and then polls the GPU continously. Luckily MW is very regular, so this uses some more CPU but otherwise works very well. Now I can use all CPU cores again.

Someone could test this by either freeing one core or by using the new swan_sync=1 option to extract maximum GPu performance. If in the latter case the difference is still there I suggest it has nothing to do with the OS, setup, services Aero etc. but rather with the driver itself. After all, the driver is still responsible for compiling the high level shader language code into machine code for the GPU - a rather complex process.

MrS
____________
Scanning for our furry friends since Jan 2002

I think the problem with W7 is the driver (or other Vista/Win7 drivers that change GPU/Motherboard power saving settings). NVidia dont even support Win 2003 server or Win 2008 server! A poor show by NVidia.

swan_sync=0 enables the use of one CPU (actually, I think it removes paramiters that restrict the CPU usage)?!?

- Paramiter is tested and works better with 6.10.73 and 6.10.72 tasks.

Judging by [rl=http://www.gpugrid.net/forum_thread.php?id=2129&nowrap=true#16920]this[/url] post the CPU polling is not the cause. So it's just plain simply the different driver?

MrS
____________
Scanning for our furry friends since Jan 2002

Ups.. thanks!

BTW: MW and CC run so well (99%) because they're so simple - it's just the same small loop over and over again, just with different numbers. Massively parallel and not really constrained by memory bandwidth. It's the absolute best case scenario for GPU code. It's not as straight forward for GPU-Grid, e.g. you have to finish one time step before you can execute the next one and you're more memory bandwidth bound. I don't know how this utilization is calculated - does it count "waiting for memory" as busy (because they SM can't do anything else right now) or is it idle time (because it's not doing anything useful)?

ATI isn't a magical bullet. Just look how hard it's been to make GPU-Grid on a current gen ATI run half as fast as a previous gen nVidia.

MrS
____________
Scanning for our furry friends since Jan 2002