Anyone else using CUDA 3.0 dlls?

Message boards : Number crunching : Anyone else using CUDA 3.0 dlls?
Message board moderation

To post messages, you must log in.

1 · 2 · 3 · 4 · Next

AuthorMessage
Profile Careface

Send message
Joined: 6 Jun 03
Posts: 128
Credit: 16,561,684
RAC: 0
New Zealand
Message 951053 - Posted: 30 Nov 2009, 12:44:16 UTC
Last modified: 30 Nov 2009, 12:45:16 UTC

Just started using them ~12hours ago (along with nvidia 195.62 WHQL drivers) and it seems to be giving me ~50-70sec decrease in WU times over CUDA 2.3 dlls (over ~0.43 AR, currently the only lot of WU I have..From ~690sec -> ~625sec)

Anyone else playing around with them? They're currently in beta, but they seem to be working well enough :)
ID: 951053 · Report as offensive
Crun-chi
Volunteer tester
Avatar

Send message
Joined: 3 Apr 99
Posts: 174
Credit: 3,037,232
RAC: 0
Croatia
Message 951058 - Posted: 30 Nov 2009, 12:53:51 UTC - in response to Message 951053.  

To use cuda 3.0 you must use 3.0 versions of cuda dll`s. And since those dll`s are not released yet, you can use 2.3 :)
WHQL drivers are support CUDA 3.0 but no one has it :)

I am cruncher :)
I LOVE SETI BOINC :)
ID: 951058 · Report as offensive
Profile Careface

Send message
Joined: 6 Jun 03
Posts: 128
Credit: 16,561,684
RAC: 0
New Zealand
Message 951062 - Posted: 30 Nov 2009, 13:08:00 UTC - in response to Message 951058.  
Last modified: 30 Nov 2009, 13:10:16 UTC

To use cuda 3.0 you must use 3.0 versions of cuda dll`s. And since those dll`s are not released yet, you can use 2.3 :)
WHQL drivers are support CUDA 3.0 but no one has it :)



30/11/2009 1:31:06 p.m. NVIDIA GPU 0: GeForce GTX 260 (driver version 19562, CUDA version 3000, compute capability 1.3, 896MB, 653 GFLOPS peak)


I'm fairly certain I'm using the 3.0 dlls. The CUDA 3.0 SDK toolkit was released beta on November 5th, just didn't realise it until today. I installed the toolkit, and moved the updated cufft.dll and cudart.dll (which have been renamed, so I updated my app_info.xml) to the setifolder, and bang.
ID: 951062 · Report as offensive
woodenboatguy

Send message
Joined: 10 Nov 00
Posts: 368
Credit: 3,969,364
RAC: 0
Canada
Message 951064 - Posted: 30 Nov 2009, 13:12:05 UTC

Any insight into when they might be released generally?

Regards,
ID: 951064 · Report as offensive
Profile Careface

Send message
Joined: 6 Jun 03
Posts: 128
Credit: 16,561,684
RAC: 0
New Zealand
Message 951065 - Posted: 30 Nov 2009, 13:15:59 UTC - in response to Message 951064.  

Any insight into when they might be released generally?

Regards,


Not a clue, but I am assuming http://forums.nvidia.com/index.php?showtopic=149959 that this is what people are referring to when they are talking about CUDA dlls. cufft.dll and cudart.dll, no?

If so, that toolkit has in the /bin/ folder two dlls called cufft32_30_8.dll and cudart32_30_8.dll, which are 6.57MB and 276kb respectively. Similar to the sizes of CUDA 2.3dlls cufft.dll/cudart.dll

Or am I missing something? o_O
ID: 951065 · Report as offensive
Crun-chi
Volunteer tester
Avatar

Send message
Joined: 3 Apr 99
Posts: 174
Credit: 3,037,232
RAC: 0
Croatia
Message 951072 - Posted: 30 Nov 2009, 14:02:32 UTC - in response to Message 951062.  

Dont work for me: computation error.
Back to cuda 2.3 dlls
I am cruncher :)
I LOVE SETI BOINC :)
ID: 951072 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 951291 - Posted: 1 Dec 2009, 6:01:30 UTC
Last modified: 1 Dec 2009, 6:02:00 UTC


~ 10 % speed up?

I don't think so.


The opt. crew made a test and saw ~ 2 % speed up, if any.

ID: 951291 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 951292 - Posted: 1 Dec 2009, 6:09:19 UTC
Last modified: 1 Dec 2009, 6:19:05 UTC


If I can quote Jason G from the opt. crew..

I dispute that the changes would amount to 2% consistently ... since Cuda tasks take such short time this 2% can be other natural variation. The tests were done with short synthetic tests, which typically vary as much as +/- 5% depending on machine state & how long windows has been running. There just aren't any changes in cuda libraries we would use until fermi boards are available.

To now I'm crunching with nVIDIA_driver_190.38 and CUDA_V2.3 .

ID: 951292 · Report as offensive
Profile Careface

Send message
Joined: 6 Jun 03
Posts: 128
Credit: 16,561,684
RAC: 0
New Zealand
Message 951307 - Posted: 1 Dec 2009, 7:39:39 UTC - in response to Message 951292.  


~ 10 % speed up?

I don't think so.


The opt. crew made a test and saw ~ 2 % speed up, if any.


Well, for me on almost identical AR WU (~0.001-0.005 variation) I'm seeing these consistent drops in crunch time, so.. For what it's worth, my setup isn't exactly typical of a CUDA cruncher, so it might account for my seeing a difference when others don't.


If I can quote Jason G from the opt. crew..

I dispute that the changes would amount to 2% consistently ... since Cuda tasks take such short time this 2% can be other natural variation. The tests were done with short synthetic tests, which typically vary as much as +/- 5% depending on machine state & how long windows has been running. There just aren't any changes in cuda libraries we would use until fermi boards are available.

To now I'm crunching with nVIDIA_driver_190.38 and CUDA_V2.3 .


If it was simply natural variation of WU times, it seems statistically highly unlikely that all post-update dlls would show extreme ~10% variation, but again, synthetic benchmarks are fine, but they do not account for all real world scenarios. Perhaps I just lucked out and got to see an increase in performance when others didn't.

Having said that, this thread isn't about arguing whether or not my observations are wrong, but whether or not you are using the dlls. If you haven't used them, how do you know that there isn't a 10% increase in performance on your system? o_O because the synthetic tests on someone else's system don't say so? o_O
ID: 951307 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 951311 - Posted: 1 Dec 2009, 8:00:46 UTC
Last modified: 1 Dec 2009, 8:11:11 UTC


It wasn't meant offensive.


You used before also nVIDIA_driver_195.62 ?
And only after updating from CUDA_V2.3 to CUDA_V3.0_BETA you saw a speed up?

If not, then I would guess the update from nVIDIA_driver_191.x to .._195.x gave the speed up.

The opt. crew (and maybe others with coding experiences) should know, if the update to CUDA_V3.0(_BETA) would be give a speed up with the current available GPUs.

For now I'll not use BETA software, I wouldn't like to damage my (expensive) hardware.
But maybe..
Nobody knows.. maybe in future yes..


BTW.
For to make a real test, the same WU should be used with different settings/software versions.
From my experiences the same AR (of different WUs) can vary ~ 5 % in calculation time.

ID: 951311 · Report as offensive
Profile Careface

Send message
Joined: 6 Jun 03
Posts: 128
Credit: 16,561,684
RAC: 0
New Zealand
Message 951334 - Posted: 1 Dec 2009, 10:11:56 UTC - in response to Message 951311.  


It wasn't meant offensive.


You used before also nVIDIA_driver_195.62 ?
And only after updating from CUDA_V2.3 to CUDA_V3.0_BETA you saw a speed up?

If not, then I would guess the update from nVIDIA_driver_191.x to .._195.x gave the speed up.

The opt. crew (and maybe others with coding experiences) should know, if the update to CUDA_V3.0(_BETA) would be give a speed up with the current available GPUs.

For now I'll not use BETA software, I wouldn't like to damage my (expensive) hardware.
But maybe..
Nobody knows.. maybe in future yes..


BTW.
For to make a real test, the same WU should be used with different settings/software versions.
From my experiences the same AR (of different WUs) can vary ~ 5 % in calculation time.


Sorry I took it offensively - it's often difficult to show feelings via text... I had been using 195.55 Beta drivers for a while, but found them to be unstable so I rolledback to 191.07. I didn't see any noticable difference in SETI, but I do admit it could very well have contributed to the performance increase.

I guess either way, whether or not its the drivers or the dlls, I'm seeing a nice change which I won't complain about :)

I'm just surprised that other users haven't spoken up about these dlls, even if it's only 2-5% over new dlls and drivers, isn't that what the number crunching forums is for? :)
ID: 951334 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 951337 - Posted: 1 Dec 2009, 10:29:21 UTC


Yes.. sure.. for this is the NC subforum..

But OTOH sometimes.. Q&A/CUDA or NC subforum..
In the Q&A/CUDA subforum I posted that there are the new driver and CUDA_BETA available.

Sometimes CUDA related threads will be moved from NC- to the Q&A/CUDA subforum.
So sometimes I don't know where I should post..


I don't know if I should 'risk' to take the BETA software.
~ 2 % speed up, would mean at my GPU cruncher + ~ 1,000 RAC.

But if this BETA(_1 - AFAIK) is too buggy, I don't want to 'damage' hard-/software.

Or I couldn't damage my GPUs or the OS/software?
BETA isn't final.

ID: 951337 · Report as offensive
Crun-chi
Volunteer tester
Avatar

Send message
Joined: 3 Apr 99
Posts: 174
Credit: 3,037,232
RAC: 0
Croatia
Message 951341 - Posted: 1 Dec 2009, 11:08:35 UTC - in response to Message 951337.  

GTS 260 is cuda 1.3, and you use cuda 3.0. I cannot use cuda 3.0 because I get computer errors.
So is it possible that new cuda works only on cuda 1.3 cards?

I am cruncher :)
I LOVE SETI BOINC :)
ID: 951341 · Report as offensive
Profile -= Vyper =-
Volunteer tester
Avatar

Send message
Joined: 5 Sep 99
Posts: 1652
Credit: 1,065,191,981
RAC: 2,537
Sweden
Message 951347 - Posted: 1 Dec 2009, 11:53:22 UTC - in response to Message 951341.  

Ahh of course, there we could have it.
I tried them on my 8800GTX and got alot of errors.
Will try them on my main rig as soon as i can download work again.

Kind regards Vyper

_________________________________________________________________________
Addicted to SETI crunching!
Founder of GPU Users Group
ID: 951347 · Report as offensive
Crun-chi
Volunteer tester
Avatar

Send message
Joined: 3 Apr 99
Posts: 174
Credit: 3,037,232
RAC: 0
Croatia
Message 951348 - Posted: 1 Dec 2009, 11:57:47 UTC - in response to Message 951347.  

9600 also gives error, as my 9800, and now your 8800 also give error, so that could be factor: cuda capability of card.
I am cruncher :)
I LOVE SETI BOINC :)
ID: 951348 · Report as offensive
Profile jason_gee
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 24 Nov 06
Posts: 7489
Credit: 91,093,184
RAC: 0
Australia
Message 951398 - Posted: 1 Dec 2009, 15:12:22 UTC - in response to Message 951307.  
Last modified: 1 Dec 2009, 15:22:51 UTC


~ 10 % speed up?

I don't think so.


The opt. crew made a test and saw ~ 2 % speed up, if any.


Well, for me on almost identical AR WU (~0.001-0.005 variation) I'm seeing these consistent drops in crunch time, so.. For what it's worth, my setup isn't exactly typical of a CUDA cruncher, so it might account for my seeing a difference when others don't.


If I can quote Jason G from the opt. crew..

I dispute that the changes would amount to 2% consistently ... since Cuda tasks take such short time this 2% can be other natural variation. The tests were done with short synthetic tests, which typically vary as much as +/- 5% depending on machine state & how long windows has been running. There just aren't any changes in cuda libraries we would use until fermi boards are available.

To now I'm crunching with nVIDIA_driver_190.38 and CUDA_V2.3 .


If it was simply natural variation of WU times, it seems statistically highly unlikely that all post-update dlls would show extreme ~10% variation, but again, synthetic benchmarks are fine, but they do not account for all real world scenarios. Perhaps I just lucked out and got to see an increase in performance when others didn't.

Having said that, this thread isn't about arguing whether or not my observations are wrong, but whether or not you are using the dlls. If you haven't used them, how do you know that there isn't a 10% increase in performance on your system? o_O because the synthetic tests on someone else's system don't say so? o_O


Allow me to clarify *my* position as best I can, so as not to have one small point out of a lengthy conversation taken out of context.

So far, when nVidia has released updates to the Cuda toolset, the improvements have been a phased refinement and extension clearly documented in release notes / readmes. In the case of 2.2 we saw a major effort at stability & bug fixes, which could arguably be called, IMO, the first really really useful SDK.

Next with 2.3 we saw a major effort specifically targeted at improvement to FFT performance ( Fast fourier transforms account for around 50%+ of multibeam processing depending on angle range)... while these present a fairly radical speed improvement for the functions we use, Cuda 2.3 capable drivers were not available for all windows platforms supported at Lunatics at the time, and when the current installer was made in particular, the 2.3 DLLs were still in beta (which is NOT public release BTW. nVidia Beta access is under closed NDA (registered developer program). Getting beta libraries yourself somehow through the backdoor, or from nVidia directly through their forums isn't in our consideration, since *we* agree to terms to have access... So nVidia & you can pretty much do what you want .. *we* can't go handing out libraries or beta DLLs etc.

Along comes v3.0 to beta, which the release notes clearly indicate is a major revision extending to new, as yet unreleased hardware.

*we* of course would like to fully support this new hardware when it materialises, and builds will exist to take advantage of certain hardware & library improvements not available yet. None of the functions used in our builds show evidence of enhancement. (This doesn't discount certain performance or bug fixes directed at specific cards)

Now drivers on the other hand, undergo constant stability and performance improvements that do affect either specific cards, through to generally applicable issues. I have little doubt that the drivers have been steadily/consistently improving (from experience alone).

In those respects, it's important to differentiate between what are generally applicable, repeatable & noticeable gains across the board, and fixes for specific hardware.

In other words, by all means experiment and find what works best for yourselves, but please don't:
(a) expect special builds just for you ... (speaking for myself really) we don't have the time to tailor to each individuals needs/wants. This is why we make source code available, so YOU, can use the toolkit you have and tailor a build for yourself.
(b) expect us to release software based on beta libraries. beta code is at all times to be considered unstable, and so not usable for release, and most likely covered under NDA we have agreed to. (what nVidia or you do on the other hand is between you and them).

Adding to that one more thing for clarification. If, adding straight new DLLs to existing Lunatic's builds, you happen to be seeing a 10% performance boost at all angles ranges, on a few different cards, then that is worthy of *us* taking a deeper look.

It's difficult to not get overly excited when nVidia release a new toolkit, after the great benefit to us from 2.3, just don't expect that every time. 2.3 addressed some pretty well known deficiences to acheive that leap.

Let's hope the upcoming new hardware releases are up to scratch, and that us poor developers can get our hands on a reasonably priced model soon enough.

Best Regards, Jason
"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.
ID: 951398 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 951496 - Posted: 1 Dec 2009, 23:06:29 UTC - in response to Message 951062.  



30/11/2009 1:31:06 p.m. NVIDIA GPU 0: GeForce GTX 260 (driver version 19562, CUDA version 3000, compute capability 1.3, 896MB, 653 GFLOPS peak)

This string alone doesn't mean that CUDA 3.0 is in use.
It just means that CUDA 3.0 supported by installed video-driver, no more.
I see such line for my host now, after driver update, but I'm absolutely sure I use CUDA 2.3 DLLs with anonymous platform.
Moreover, BOINC can't determine what DLLs/CUDA version is used by what project.
Few GPU projects can easely use different CUDA DLLs with their applications.

To use CUDA 3.0 one need to put them explicitely as replacement of CUDA 2.3 DLLs into SETI project folder.
(Can't comment performance change 2.3 vs 3.0 still).
ID: 951496 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 951497 - Posted: 1 Dec 2009, 23:08:27 UTC - in response to Message 951341.  


So is it possible that new cuda works only on cuda 1.3 cards?

Highly unlikely. But you can check it on nVidia's pages...
ID: 951497 · Report as offensive
Profile perryjay
Volunteer tester
Avatar

Send message
Joined: 20 Aug 02
Posts: 3377
Credit: 20,676,751
RAC: 0
United States
Message 951506 - Posted: 1 Dec 2009, 23:32:09 UTC - in response to Message 951497.  

Thanks for pointing that out Raistmer I hadn't noticed mine says version 3000 too.

12/1/2009 5:02:31 PM NVIDIA GPU 0: GeForce 9500 GT (driver version 19562, CUDA version 3000, compute capability 1.1, 1024MB, 118 GFLOPS peak)

Must be because of the drivers because I'm using the 2.3 DLLs.


PROUD MEMBER OF Team Starfire World BOINC
ID: 951506 · Report as offensive
_heinz
Volunteer tester

Send message
Joined: 25 Feb 05
Posts: 744
Credit: 5,539,270
RAC: 0
France
Message 951508 - Posted: 1 Dec 2009, 23:41:04 UTC

new NVIDIA driver 19562 CUDA version 3000 is now available for download...
01.12.2009 22:29:27 NVIDIA GPU 0: ION (driver version 19562, CUDA version 3000, compute capability 1.1, 256MB, 35 GFLOPS peak)

this answers a lot of questions.
ID: 951508 · Report as offensive
1 · 2 · 3 · 4 · Next

Message boards : Number crunching : Anyone else using CUDA 3.0 dlls?


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.