That is almost 60% faster for the IBUCH tasks under Win7 using the 6.11 app.
[...]
roundup, I think it is worth a try; as long as you get plenty of IBUCH tasks, it should be a good improvement. If you get an even spread of tasks (with 10%, 30% and 60% improvement), it should cover the drivers losses (as long as it was not over 33%)!

I am really happy to have installed the new driver. Thanks for the advice, skgiven!
This computer never completed a 6803 credit task (TONI_CAPBIND) under Vista/7 in about 10k seconds before:
http://www.gpugrid.net/result.php?resultid=2700834.
Obviously driver 258.96 running CUDA3.1 tasks is much faster under Vista/7 than driver 257.21 running CUDA3.0 units - the difference is about 25% for the TONI tasks. With the former 257 driver it took an average of 16k seconds for a 5,965 credits task. Example (IBUCH): http://www.gpugrid.net/workunit.php?wuid=1713213.

The latest 258 driver with CUDA3.1 is a big step forward for the Win7 users.

Obviously driver 258.96 running CUDA3.1 tasks is much faster under Vista/7 than driver 257.21 running CUDA3.0 units - the difference is about 25% for the TONI tasks. With the former 257 driver it took an average of 16k seconds for a 5,965 credits task. Example (IBUCH): http://www.gpugrid.net/workunit.php?wuid=1713213.

I think you need to compare similar WUs. This computer shows 5 of the 6803 credit TONI_CAPBIND WUs. The 2 run under CUDA30 had times of 12,994 and 13905. There were 3 run under CUDA31, 2 successes with times of 10,330 and 10,076, and 1 failure with a time of 15,897. In the best of worlds nowhere near 25% and if you count the failure actually slower for CUDA31.

Obviously driver 258.96 running CUDA3.1 tasks is much faster under Vista/7 than driver 257.21 running CUDA3.0 units - the difference is about 25% for the TONI tasks. With the former 257 driver it took an average of 16k seconds for a 5,965 credits task. Example (IBUCH): http://www.gpugrid.net/workunit.php?wuid=1713213.

I think you need to compare similar WUs. This computer shows 5 of the 6803 credit TONI_CAPBIND WUs. The 2 run under CUDA30 had times of 12,994 and 13905. There were 3 run under CUDA31, 2 successes with times of 10,330 and 10,076, and 1 failure with a time of 15,897. In the best of worlds nowhere near 25% and if you count the failure actually slower for CUDA31.

I know why there was a failure: My fault, when I tested with different BOINC and environmental settings.
Also the HIVPR units are much quicker: Now 8,759.88 seconds. I never delivered a HIVPR unit with less than 11k seconds before, not even on a different (and much faster) PC that also is equipped with a GTX470.

I know why there was a failure: My fault, when I tested with different BOINC and environmental settings.
Also the HIVPR units are much quicker: Now 8,759.88 seconds. I never delivered a HIVPR unit with less than 11k seconds before, not even on a different (and much faster) PC that also is equipped with a GTX470.

That is faster. Just wondering, what BOINC & environmental settings are you referring to? Also did you use swan_sync for both CUDA30 & CUDA31?

I saw these two KASHIF_HIVPR tasks, one using CUDA3.0 and then using CUDA3.1 (with the latest molecular dynamics app; v6.11). They ran on the same Vista system. The v6.11 app was 44% faster:

2616205 1663394 4 Jul 2010 11:22:11 UTC 9 Jul 2010 10:27:45 UTC Completed and validated 12,636.57 12,351.77 4,428.01 4,428.01 ACEMD2: GPU molecular dynamics v6.05 (cuda30)

2701801 1713122 21 Jul 2010 10:42:25 UTC 21 Jul 2010 14:14:34 UTC Completed and validated 8,759.88 8,882.59 4,428.01 6,642.02 ACEMD2: GPU molecular dynamics v6.11 (cuda31)

I ran a similar KASHIF_HIVPR task on my GTX470, which took 7417sec (257.21). My task ran about 18% faster than yours, both using v6.11, but my system is overclocked to 704MHz (about 16% faster than stock), and I also freed up 2 cores on my i7-920.

Roundup, is your GTX470 at stock?
I'm asking because I am using XP x86!

I have seen some very significant improvement in a few specific WUs.

However I thought I would post results from one of the seemingly least efficient WUs I have run: IBUCH_freebind_pYEEI_100706. This one only pushes my GPU usage into the 70 something percent.

2670817 16,609.06 16,413.48 7,954.42 11,931.63 ACEMD2: GPU molecular dynamics v6.05 (cuda30) 13.284 ms/step

2701432 15,484.38 15,454.58 7,954.42 11,931.63 ACEMD2: GPU molecular dynamics v6.11 (cuda31) 12.385 ms/step

That's approximately a 7% improvement.

I agree that different work types are showing different performances, but the important thing is that they have all improved. No Fermi tasks perform worse using the latest 6.11 app. The latest improvements have allowed the GTX460 cards to work, and brought about a big improvement for Vista and W7 users (who needed it the most).
Like you, I have a GTX470 and use XP. Yesterday and this morning I ran tasks using the older 25721 driver, as you are doing. I have now changed to the latest driver, just on the off chance it brings an improvement my way. I dont think it will (just for Vista and W7) but I will try it and report back.

I saw these two KASHIF_HIVPR tasks, one using CUDA3.0 and then using CUDA3.1 (with the latest molecular dynamics app; v6.11). They ran on the same Vista system. The v6.11 app was 44% faster:

[...]
My task ran about 18% faster than yours, both using v6.11, but my system is overclocked to 704MHz (about 16% faster than stock), and I also freed up 2 cores on my i7-920.

Roundup, is your GTX470 at stock?
I'm asking because I am using XP x86!

The GTX470 is at 702/1708/1404, SWAN_SYNC is set. I only freed up one core on the i7-920 @ stock.
The 44% advantage of CUDA3.1 over CUDA3.0 is amazing.

I think your 18% advantage is caused by 2 instead of 1 freed up core and XP instead of Vista.
The disadvantage of Vista/7 has become smaller, but is still there.

I've got another "...cuda31.exe has stopped working" error (during watching some HD video), a failed task, and a task cancelled by server. This last thing is a novelty for me.

On my GTX470 (XP), I found no performance change using the latest driver (25896), instead of the older 25721 driver,
[edit]
but I did find that the 6.11 app brought an average 11% improvement from the 6.05 app:

- IBUCH_freebind_pYEEI . . . . 9.85% Faster
- IBUCH_101b_pYEEI . . . . . . 10.25%
- TONI_CAPBINDsp1 . . . . . . . 12.00%
- TONI_CAPBINDsp2 . . . . . . . 12.18%
- KASHIF_HIVPR_auto_spawn 10.48%
[/edit]

roundup, I think having 2 cores free expidites the task by about a further 3% but no more than about 8%, so a GTX470 may only be 10 to 15% slower while on Vista/Win7 than it would be on XP. More task types required to be sure about this generalization, and there are task variations.
Going by your tasks, on Vista, moving to the latest driver and running the 6.11 app you saw these improvements:

- IBUCH_201b_pYEEI, 57% faster
- KASHIF_HIVPR, . . . 39% faster
- TONI_CAPBINDsp2, 33% faster
- TONI_CAPBINDsp1, 26% faster

Looks like, my 480 cards running on Win 7, are around 30% to 40% faster on acemd2 611 compared to 605, which means that it is, on average, slightly faster than a 285 card running on win xp. This is a good improvement, but 480 should really be twice as fast as 285, since it has double the cores. So the drivers for these cards, still need to be updated.

Looks like, my 480 cards running on Win 7, are around 30% to 40% faster on acemd2 611 compared to 605, which means that it is, on average, slightly faster than a 285 card running on win xp. This is a good improvement, but 480 should really be twice as fast as 285, since it has double the cores. So the drivers for these cards, still need to be updated.

Apples to oranges
You can't really make a comparison to a new card running on Win7 to an older card running on WinXP

The drivers on Win7 don't need to be "updated" M$ needs to abandon WDDM which has unacceptable overhead that directly and significantly affects performance.

IMHO

Looks like, my 480 cards running on Win 7, are around 30% to 40% faster on acemd2 611 compared to 605, which means that it is, on average, slightly faster than a 285 card running on win xp. This is a good improvement, but 480 should really be twice as fast as 285, since it has double the cores. So the drivers for these cards, still need to be updated.

Apples to oranges
You can't really make a comparison to a new card running on Win7 to an older card running on WinXP

The drivers on Win7 don't need to be "updated" M$ needs to abandon WDDM which has unacceptable overhead that directly and significantly affects performance.

IMHO

One question, just to make it clear:
If a HW-monitor says, GPU is running at ~95% on Win7, does that mean, that it is as efficient as running at 95% on XP ? Currantly I do not own a Fermi-card (only an GTX260/192), but I plan to buy a GTX460 and it would make only little problems to swich back to XP.

Regards,
Alexander

No. Win7 is not as efficient as XP. Under Win7 the 95% total includes the Win7 Driver overhead (perhaps 10 or 15%). Win7 is still slower than WinXP due to the Driver Architecture.
The last GPUGrid application, along with the latest driver, resulted in a speed bump for the first Fermi cards (GF100) and enabled the second Fermi cards (GF104) to work. Prior to these updates the GF100 cards were almost half as slow on Vista and Win7 than on XP.

Uups!!
And what about Kubuntu? On one of my systems I could run the required apps in a VM under linux.

Regards,
Alexander

One question, just to make it clear:
If a HW-monitor says, GPU is running at ~95% on Win7, does that mean, that it is as efficient as running at 95% on XP ? Currantly I do not own a Fermi-card (only an GTX260/192), but I plan to buy a GTX460 and it would make only little problems to swich back to XP.

Regards,
Alexander

In one word:
No.

I have WinXP, Vista, and Win7 on the same hardware. I started GPUgrid on Win7, and the GPU monitoring was showing very high percentage, but the tasks took almost twice as much time than on (other's) WinXP. Then I learned about the WDDM overhead, therefore I switched 'back' to WinXP. The GPU monitoring still showing the same high GPU usage, but the tasks run much faster. And one dedicated core with the SWAN_SYNC=0 environmental value setting made the tasks run a little more (about 10-15%) faster.
Switching back to XP is recommended for the sake of crunching efficiency. You can even create a multi-boot environment: WinXP for crunching, and Win7 for gaming. Or you can have a GPUgrid running under linux on a USB stick.

Dont boot to XP first and then to a Linux Virtual Machine.
First of all you can't use XP x86 as the Linux app is X64 (the Windows app is x86). Although you could do it from XP x64 you would have the overhead of XP, the Virtulisation software and Linux, and three sets of potential problems. Better to use either XP or Linux.
If you must use Vista or Win7 then Disk/USB booting to Linux (FatPuppy) is an option.

Thanks a lot for your answers.

My main system has to run under an 64Bit OS due to mem requirements, and it must be windows. Since XP64 is no option, I have to use Win7. (Once upon Vista I had a working system ;-)

My other system is dual boot XP/Win7. I can check the speed difference there. Its not necessary to run Win7 all days.
My idea was to install Kubuntu and run the two most used windows-apps in a VM under linux. These apps are tested and work fine in a VM. But will that bring an remarkable increase in speed?

Regards,
Alexander

My other system is dual boot XP/Win7. I can check the speed difference there.

I tested it overnight. Since GPUGRID fails on my GTX260/192, I had to test it with Einstein and Milkyway. And YES, you're absolutely right, its pretty much faster! I don't have exact numbers, but it seems to be in the range of 20%.

Alexander

I tested it overnight. Since GPUGRID fails on my GTX260/192, I had to test it with Einstein and Milkyway. And YES, you're absolutely right, its pretty much faster! I don't have exact numbers, but it seems to be in the range of 20%.

You have a speed difference between WinXP and Win7 in MilkyWay? At least with the ATI cards there's no difference at all, not even one second.