Found this on Nvidia's forums:

No, there is no way around it. XP used XPDM, while Vista and Windows 7 use WDDM. WDDM is radically different, and basically 'forces' the graphics drivers to do things a certain way to ensure the stability of the operating system. However, some of the extra code also causes some kernel call overhead, which you really can't do much about, other than to try to figure out a way to call your kernels less frequently. WDDM shouldn't slow the kernel itself down though.

Note that 'kernel' in this context is not the typical OS kernel that many of us are familiar with. A 'kernel', in Nvidia's nomenclature, is the section of C code that is dispatched to run on the GPU.

Of interest to developers is that it might be adventageous to create larger blocks for the GPU to execute in order to avoid the overhead in Vista and Windows 7.

Mike

---

Want to find one of the largest known primes? Try PrimeGrid. Or help cure disease at WCG.

I accept that Vista approaches tasks differently than XP, but for me Vista is faster on the whole (IF YOU CONFIGURE IT CORRECTLY), is more functional and significantly more financially viable.
I have a Phenom 9750 and with Vista, the system boots and shuts down much quicker than it did with XP, it is almost silent, and that is while running GPUGRID non-stop.
I set the Power Options to Power Saver and Vista configured the system to operate at lower frequencies. This has reduced my overheads considerably and running GPUGrid makes no impact!
When I run Boinc for other projects the impact is much higher in terms of Watts, so this one is a winner for me with Vista!

Yup, yup. I prefer Vista also; I wasn't suggesting that XP (or Vista for that matter) was the better version of the operating system.

There had been reports that for some reason, GPUGRID appeared to be running slower on Vista than on XP (on the same hardware), and I was posting a link to an explanation of why that would be. Unless a given application is able to take into account the higher overhead, GPU tasks are going to be slower on Vista.

Mike

Last autumn it seemed like we were seeing faster crunching times on Vista, possibly due to better optimizations in the new (radically different) driver.

The kernel overhead is something game developers really have to consider, as they're working "closely" with te GPU. GPU-Grid, however, is working "loosely" with the GPU, i.e. the interaction (sending a new kernel) happens seldomly. So they're actually already doing what you suggest :)
I suppose each "step" in each WU involves a new kernel call, or maybe some data is just being udated. This happens every 20 - 500 ms, depending on GPU speed. So I suggest that speed differences between the OSes are likely not caused by this overhead, but rather by differences in the driver and how it handles CUDA.

MrS

---

Scanning for our furry friends since Jan 2002

There had been reports that for some reason, GPUGRID appeared to be running slower on Vista than on XP (on the same hardware), and I was posting a link to an explanation of why that would be. Unless a given application is able to take into account the higher overhead, GPU tasks are going to be slower on Vista.

Didn´t run both cards (same OC 666/1512/1150 MHz, different SP) long enough to get a good base to compare hosts but as an example:

* 29460 > GTX 260/192 SP: 42-43 ms time/step, CPU-Time > 4000 s, Windows7/64, X4 810 @2,6GHz
* 23101 > GTX 260/216 SP: 33-34 ms time/step, CPU-Time > max. 1500 s, WinXP/32, Q6600 @3,0GHz (9800GX2 before, don´t be confused)

BOINC mostly serves same projects on both hosts: QMC@home, Spinhenge etc. The differences can not be explained in the number of shader processors at all, in my opinion.

Could you upgrade the driver on the Win7 box to the 185 series? A temperature increase has been observed with these under Vista, which should mean somewhat faster processing.

EDIT: and I moved my 9800GTX+ from my XP32 machine to a Vista 64 machine 1 or 2 months ago. Direct comparisons were difficult, but I was at ~12h per task and am at about 11h now.
So it's not that Vista would be ~30% slower in general. But still i'd like to know more, since you're the second to report such a difference for GT200-based cards.

MrS

---

Scanning for our furry friends since Jan 2002

Could you upgrade the driver on the Win7 box to the 185 series?

Yes I installed now 185.68 (Beta) for Vista/64, effects will be seen with the next new workunit, the current one (580445) is finished ~50%, perhaps there is in fact a change, thanks for your advice.

(I used in a final step the official driver for some tests relating to HybridPower, but because the system is now almost using the GPU it´s not necessary any more.)

Yes I installed now 185.68 (Beta) for Vista/64, ...

... unfortunately caused a compute error.

Another error (582098), switched back to 181.71.

ETA, he's going faster in XP because that machine has the faster card. 216 core vs 192.

ETA - What temperature is your 9800GTX+ running at ?

Regards
Zy

Mhh, sorry or wasting your time with that driver. Could be that the Vista drivers are not always so surprisingly compatible with Win 7.

@Jeremy: his cards run at the same clocks, so the only difference are the 12.5% more shaders, so his XP setup should be about 12.5% faster. However, he's seeing about a 27% difference:

X1900AIW wrote:

Didn´t run both cards (same OC 666/1512/1150 MHz, different SP) long enough to get a good base to compare hosts but as an example:

* 29460 > GTX 260/192 SP: 42-43 ms time/step, CPU-Time > 4000 s, Windows7/64, X4 810 @2,6GHz
* 23101 > GTX 260/216 SP: 33-34 ms time/step, CPU-Time > max. 1500 s, WinXP/32, Q6600 @3,0GHz (9800GX2 before, don´t be confused)

BOINC mostly serves same projects on both hosts: QMC@home, Spinhenge etc. The differences can not be explained in the number of shader processors at all, in my opinion.

@Zydor: about 55°C with 2 120mm fans at ~500 rpm on an Acceleron S1 Rev 2. Why?

MrS

---

Scanning for our furry friends since Jan 2002

I wonder if the driver differences could explain the rest of the difference (beyond the 12.5% more due to 216 vs. 192 shaders)? The Win7 box is using 181.71. The XP box is using 182.06. Anyone know any particular driver enhancements in the latter that might increase performance slightly?

Not yet for win7 but for vista and xp there are some beta's released which should give a speed increase on these newer cards.

The versions i found are version 185.68 for them but i am warning people when use them make sure no units are running on projects because a change on loss of everything is a BIG YES.

Again these versions seems not to run on win7 but to be honest i have not tried them on my box yet

And YES Vista is slower then XP absolute confirmed,
But you can get close to XP performance on vista if you have good drivers and tune the os to run as little as possible of these graphical toys, less is better.
Also turn off crap like indexer and other slow your system down services which are not needed for boinc ;)

But ofcourse i do that for the xp machines as well and.
Still xp is faster untill now, although i got a the idea that win7 has potential to beat XP.
My test have confirmed a faster boot then XP but the lack of speedy drivers for it make it hard to confirm that it can beat XP on projects
I guess time will tell when the RC gets released for the big public to test and hopefully M$ is not screwing it up ;)
Don't get me wrong i think we really can start testing if win7 will be faster when the big hardware builders start to release drivers for win7, untill that has happened we are stuck on what M$ has made for it ( or what was allowed by them singing)

The Win7 box is using 181.71.

Since morning 181.22 (Vista/64-WHQL) is installed.

Mhh, sorry or wasting your time with that driver.

Doesn´t matter, it´s a good investment if someone else avoids the same test cycle, the card is high overclocked (original voltages however), I can´t complain about it, should run both cards at lower clock settings, sometimes curious things happen, may be I got faster output with lower settings. After some warming up with F@H resp. "burning in" the driver I´ll test it with GPUgrid one more time.

I wonder if the driver differences could explain the rest of the difference (beyond the 12.5% more due to 216 vs. 192 shaders)? The Win7 box is using 181.71. The XP box is using 182.06. Anyone know any particular driver enhancements in the latter that might increase performance slightly?

None have been reported here.

MrS

---

Scanning for our furry friends since Jan 2002

The 185.66+ BETA drivers have been reporting to have some good effects on performance but as usual we talk about speed in games on these fora :D
But i saw that some improvements where made on CUDA as well, since this where commands only being used by g200 cards i can't confirm it.

* 29460 > GTX 260/192 SP: 42-43 ms time/step, CPU-Time > 4000 s, Windows7/64, X4 810 @2,6GHz
* 23101 > GTX 260/216 SP: 33-34 ms time/step, CPU-Time > max. 1500 s, WinXP/32, Q6600 @3,0GHz (9800GX2 before, don´t be confused)

No change in configuration since backroll to 181.22 (other BOINC projects yes > Spinhenge exclusive), it is possible to crunch faster - depending on the workunit ? 598868: credit 4352/5440

* 29460 > GTX 260/192 SP: 30 ms time/step, CPU-Time > 4700 s, Windows7/64, X4 810 @2,6GHz

is it possible to crunch faster - depending on the workunit ?

Definitely. The complexitiy of the WUs (~number of atoms) varies and therefore also the time per step. The number of steps per WU is adjusted to keep the overall length in check.

If case of this WU its name is probably "pYIpYV", but I'm not totally sure. If I compare the 2 hosts in question all of their WUs with 3946.78 claimed credit have this string in their name. And for them the times/step are pretty constant, as well the ~25% performance advantage for the XP PC with 12.5% more crunching power. The new unit you linked to is a totally different beast.

MrS

---

Scanning for our furry friends since Jan 2002

* 29460 > GTX 260/192 SP: 42-43 ms time/step, CPU-Time > 4000 s, Windows7/64, X4 810 @2,6GHz
* 23101 > GTX 260/216 SP: 33-34 ms time/step, CPU-Time > max. 1500 s, WinXP/32, Q6600 @3,0GHz (9800GX2 before, don´t be confused)

* 29460 > GTX 260/192 SP: 30 ms time/step, CPU-Time > 4700 s, Windows7/64, X4 810 @2,6GHz

To complement my comparison: 616407 (Name 81-KASHIF_HIVPR_mon_ba3-2-100-RND6706_0)

* 23101 > GTX 260/216 SP: 22.6 ms time/step, CPU-Time > 725 s, WinXP/32, Q6600 @3,0GHz