Latest drivers (NVidia and ATI)

[Jason G]

it is for free download for now ....

where's that ?  (I only see it in registered developer area, they launching in the forums again ?  )

[Devaster]

yes in the forum only ...

btw have u know that intel has bought rapidmind ?

[Jason G]

...
btw have u know that intel has bought rapidmind ?

  Ooh .. them's fighting moves, LoL.

[Devaster]

Note that for CUFFT 3.0, the layout of batched data must be side-by-side
   and not interleaved.

so it seems that we must rearrange data before fft to real parts and imaginary parts in rhe same array and after this doing it back ....

[Jason G]

Note that for CUFFT 3.0, the layout of batched data must be side-by-side
   and not interleaved.

so it seems that we must rearrange data before fft to real parts and imaginary parts in rhe same array and after this doing it back ....

Well AFAICT without bothering to look much further, they've simply added stride parameters, which would just require examining the data arrangement & setting the strides for the appropriate format you want.  No doubt that makes some more advanced multidimensional FFTs easier to implement.  For our purposes it'll probably just be a minor annoyance, to be overcome with macro definitions or wrapper functions.

[Sutaru Tsureku]

Hmm.. O.K., I saw nVIDIA released the new 195.62 driver.
For CUDA_V3.0 .

Where I could get CUDA_V3.0 , and I could update my current opt._CUDA_6.08_V12_app with CUDA_V3.0 ?

Maybe one of the opt. crew is again kind and take the two .dll's out of the 'DL pack' and publish it also in the DL area here? 
Or this CUDA_V3.0 must/should be 'modded' for SETI@home?

[Jason G]

Hi sutaru.  No significant speed advantage has been detected with 3.0.  The improvements are primarily directed at unreleased hardware.  We'll keep an eye on if the situation changes, but won't be rushing to build/release stuff for no advantage, or release SDK components under NDA (risking our nVidia licences. )

[Sutaru Tsureku]

Thanks for reply!

Normally I don't like to use BETA software.
But.. if I know (and I know ), from where I could get the CUDA_V3.0_BETA - it would run well with the current opt._CUDA_V6.08_V12_app?
Or you would say.. no, better not to use CUDA_V3.0_BETA?

If CUDA_V3.0_final will be released in future, this two .dll's will be published in the DL area here?
It need to be modded, or just 'out of the box'?

I'm current with nVIDIA_driver_190.38. CUDA_V2.3 .
WinXP 32bit, 5x OCed GTX260-216.
I would see a performance difference only with the new nVIDIA_driver_195.62 ?

AFAIK, new nVIDIA_driver are well for gaming people.
But, how it's with the only CUDA people? They profit from it also?

Sorry for my questions..

[Jason G]

Quite OK Sutaru. I would say no point risking task cache (especially with current project difficulty) until there is GT300 based card in your machine  .  2.3 addressed a known cuda speed problem and helped us directly through the functions we use so made a big speed improvement. 3.0 doesn't, it adds more stuff we don't use (Yet, LoL  ). 

Q: Do you currently run the GPUs flat out, and reschedule/kill VLARs? and leave CPU just mainly for feeding? If so I may have something worth for you to test soon, that may be a more important thing to study how it works. (after some school work is finished).

[Sutaru Tsureku]

You didn't answered all my questions. 


My GPU cruncher:
http://setiathome.berkeley.edu/show_host_detail.php?hostid=4789793
4x EVGA GTX260-216 SSC supported by AMD Phenom II X4 940 BE @ 3.0 GHz.
(nothing on CPU, 4x CUDA)

My QX6700:
http://setiathome.berkeley.edu/show_host_detail.php?hostid=5069275
1x GIGABYTE GTX260(-216) SOC supported by Intel Core2 Extreme QX6700 @ 3.14 GHz.
(4x MB, 1x CUDA)
Here I use Fred's nice prog TThrottle, which increase the CUDA priority to 'normal'.
For max. GPU performance. 

TThrottle? Small look in my profile: 
http://setiathome.berkeley.edu/view_profile.php?userid=8619447


How would be the test?
It's 'safe' for the hardware/software? 
If PM, then please over the SETI@home site. There I'm every day.. 

[Jason G]

Sorry Sutaru, normally If I don't answer something it's either because I don't know, or didn't understand the question.  In this case probably a bit of both plus many questions at once.

I don't mess with priority, so don;t have answers for those as to what's best.

The test would be for V13 hybrid which seems to work well, but some minor issues are not well understood.  It moves only the 'difficult' parts of pulsefinding to CPU, instead of rescheduling/killing VLArs.  When I have some free time to look at tests, I will PM yourself at s@h.  If it would work OK for you, then I would put it to beta soon.

[Sutaru Tsureku]

Sorry, we misunderstood.
English isn't my motherlanguage.

I posted my hardware for to show how I let run SETI@home.
I use TThrottle on the QX6700 + OCed GTX260-216 for to increase the CUDA priority (4x MB + 1x CUDA).
I don't know if it would be well (any profit) also for the GPU cruncher (4x OCed GTX260-216), for to let run 4x MB + 4x CUDA. (Because the GPUs would lose performance, and the CPU wouldn't have much for to compensate)
I calculated little bit around.. 

I have the following questions..
To hold it clear..


* You would use the current CUDA_V3.0_BETA ?
Because I see lot of people in the forum, which let run this.
Speed difference? Well for the opt._MB_6.08_V12_app ?

If ~ 2 % speed up, this would mean + ~ 1,000 RAC at my GPU cruncher.


* I would see a speed difference between:
nVIDIA_driver_190.38 + CUDA_V2.3
and
nVIDIA_driver_195.62 + CUDA_V2.3

Normally nVIDIA release new driver for the gaming community.
The only CUDA community profit also from new driver versions?
(I don't mean, new driver for new CUDA version - this is clear.. )


Thanks!

[Jason G]

....
I don't know if it would be well (any profit) also for the GPU cruncher (4x OCed GTX260-216), for to let run 4x MB + 4x CUDA. (Because the GPUs would lose performance, and the CPU wouldn't have much for to compensate

  Good.  No Cpus probably would be good for the test I have in mind.

* You would use the current CUDA_V3.0_BETA ?
Because I see lot of people in the forum, which let run this.
Speed difference? Well for the opt._MB_6.08_V12_app ?
,,,

         First initial alpha test would be an existing 2.3 build , susbsequent tweaking would likely be v3.0 Beta.

If ~ 2 % speed up, this would mean + ~ 1,000 RAC at my GPU cruncher.

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.

* I would see a speed difference between:
nVIDIA_driver_190.38 + CUDA_V2.3
and
nVIDIA_driver_195.62 + CUDA_V2.3

Normally nVIDIA release new driver for the gaming community.
The only CUDA community profit also from new driver versions?
(I don't mean, new driver for new CUDA version - this is clear.. )

For the driver change alone, I don;t think a change would be as noticable as V13 test, which should take some hard work the GPU has trouble doing away, and puts it on the CPU, with AKv8 code (like reschedule internally only the hard parts), while the fast bits still process on GPU. That should make Mid-high angle ranges ~same speed to a tiny fraction faster elpased, with little/no extra CPU use, while VLAR can process ~2x speed of AKv8 (depending on GPU & CPU ratio).

Reason for test: There are some things I need to understadn how this would work on faster cards/machines, that none of us here have the hardware to test.

Be prepared though If you might lose some cache by accident 

Jason

[msattler]

I think we have determined that the 195.62 nvidia driver is not the best thing to run on pre-Fermi cards.

Has anybody determined if there is any performance difference between the 190.62 (which is currently working very well for me) and the last pre-Fermi update 191.07?

I hate to break anything that is working, unless somebody has noticed an advantage in doing so.

[Pappa]

I think we have determined that the 195.62 nvidia driver is not the best thing to run on pre-Fermi cards.

Has anybody determined if there is any performance difference between the 190.62 (which is currently working very well for me) and the last pre-Fermi update 191.07?

I hate to break anything that is working, unless somebody has noticed an advantage in doing so.

It appears to be working well on my 250 running over in Seti Beta... It is hard to say if it really has any "real" advanatage.