Latest drivers (NVidia and ATI)

[SciManStev]

I used to use a 295, and I was told that it was a very powerful card. The card alone was generating about 8000 RAC. I kept getting -1 errors, which as I am learning may have been due to the drivers that were available at the time. I gave the card to another cruncher, and he is having better luck with it using new drivers. My two 480's have demolished anything I was getting with the dual 295, so I consider them a very good choice. Perhaps by the time a dual 480 comes out I will have the money to buy a couple of them. I can always dream.....

Steve

[Fredericx51]

Have installed/switched a GTX480 in stead of the 9800GTX+.

Quite a BIG difference, now running 3 SETI MB CUDA together, does take a lot of power to get it to
run, it has 1x 8 pin connector, 1 x6, both have to be used and be able to deliver 17A + 17A + 5- 8A via Mobo, also gets extremely hot, 85- 95C !
Had to switch PSU's to get it fed with enough power. Mobo and PSU appeared to 'light to get 2 (470) of them running on this Mobo! 

And the latest drivers, CUDA 3.1, think it can run even 4 SETI MB tasks at a time.
Now have a good reason to build another rig for the 470 ,  9800GTX+ (1.1 Compute Capabillity) doesn't work together with a 470 or 480, maybe Mobo related, but I'm not sure.

At GPUGrid you can clearly see how fast they are, running their latest WU's .

[SciManStev]

The 480's do run hot, but that's why I water cool them. I can over clock them to extremes and never see even 48C. I am currently running at 840 Mhz, with more room to play. Temps are 42 and 44C at the moment. I am running 2 wu's at a time, but perhaps experimenting when time permits can change that.

Steve

[Fredericx51]

Even on  'air -cooling' ,  the 480 can control temps, using the Fan Settings and according to NVidia's
Spec's, they are 'build' , to run quite hot.  (~100C !)
Have installed the 258.96 drivers, but it's hard to compaire anything a.t.m.  Can't even see any increase in runtime, going from 2 to 3 running parallel.  Also read, that's possible to run 4 at a time, thinking
of trying this, with stock settings.
Has anybody, with an 480 tried this and compaired runtimes, on WU's with about the same AR, f.i.
~0.4 and maybe also 0.1 or >1 ?

With SETI OnLine, it's much easier and I do hope the A/C unit failiar, hasn't done any serious damage.  Does anybody know, what the 'plan' is and when SETI is going online ?
Could be next friday, since the weekly maintenance, would have started today, anyway.......

[Ghost0210]

new Beta driver is released http://www.nvidia.co.uk/object/win7-winvista-64bit-260.63-beta-driver-uk.html version 260.63
Gonna give it a go and see what happens 

[msattler]

new Beta driver is released http://www.nvidia.co.uk/object/win7-winvista-64bit-260.63-beta-driver-uk.html version 260.63
Gonna give it a go and see what happens 

Will be interested to see what you are able to find out.

[Ghost0210]

Well the good news is it hasn't broken anything
Looks pretty similar so far, apart from a nicer looking installer, added support for the the 450 cards and some minor bug fixes doesn't look like a hell of a lot has changed
Gonna let the last 10 or so tasks I have run through then run some comparisons through using test wu's to see how these stack up against the 258.96 drivers

[msattler]

My bet is the main reason for the release was the 450s.

[Jason G]

My bet is the main reason for the release was the 450s.

... and possibly sneaking in Cuda 3.2 support before the programming tools are released.

[Ghost0210]

So have done a couple of quick bench tests with both the 260.63 and the 258.96 drivers:

260.63
Lunatics_x32f_win32_cuda30_preview.exe -verb -nog -device 0 / PG0009.wu :
Started at  : 19:39:37.231
Ended at    : 19:43:30.400
    233.103 secs Elapsed
     21.996 secs CPU time

258.96
Lunatics_x32f_win32_cuda30_preview.exe -verb -nog -device 0 / PG0009.wu :
Started at  : 19:51:29.541
Ended at    : 19:55:14.693
    224.979 secs Elapsed
     19.672 secs CPU time

258.96 faster
--------------------------
260.63
Lunatics_x32f_win32_cuda30_preview.exe -verb -nog -device 0 / PG0395.wu :
Started at  : 19:55:18.010
Ended at    : 19:56:11.311
     53.266 secs Elapsed
     14.680 secs CPU time

258.96
Lunatics_x32f_win32_cuda30_preview.exe -verb -nog -device 0 / PG0395.wu :
Started at  : 19:43:33.480
Ended at    : 19:44:41.476
     67.938 secs Elapsed
     24.274 secs CPU time

260.63 faster
--------------------------------------------
260.63
Lunatics_x32f_win32_cuda30_preview.exe -verb -nog -device 0 / PG0444.wu :
Started at  : 19:56:14.372
Ended at    : 19:57:00.696
     46.287 secs Elapsed
     14.648 secs CPU time

258.96
Lunatics_x32f_win32_cuda30_preview.exe -verb -nog -device 0 / PG0444.wu :
Started at  : 19:44:44.563
Ended at    : 19:45:41.935
     57.314 secs Elapsed
     22.261 secs CPU time

260.63 faster
--------------------------------
260.63
Lunatics_x32f_win32_cuda30_preview.exe -verb -nog -device 0 / PG1327.wu :
Started at  : 19:57:03.777
Ended at    : 19:57:41.241
     37.423 secs Elapsed
     14.196 secs CPU time

258.96
Lunatics_x32f_win32_cuda30_preview.exe -verb -nog -device 0 / PG1327.wu :
Started at  : 19:45:45.048
Ended at    : 19:46:31.642
     46.532 secs Elapsed
     19.765 secs CPU time

260.63 faster

----------------------------------------------------
Looks like overall 260.63 is faster by some margin! All wu's ran 1 at a time on my GTX465

[Frizz]

... and possibly sneaking in Cuda 3.2 support before the programming tools are released.

Looks like ...

Code: [Select]

Number of platforms: 1
  Plaform Profile: FULL_PROFILE
  Plaform Version: OpenCL 1.0 CUDA 3.2.1
  Plaform Name: NVIDIA CUDA
  Plaform Vendor: NVIDIA Corporation
  Plaform Extensions: cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_d3d9_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll


[Ghost0210]

Looks like ...

Sneaky, nothing in the release notes or published about this going in that I've been able to find
but I'm getting the same platform version of 3.2

Code: [Select]

G:\Apps\ATIStream\bin\x86_64>CLInfo.exe
Number of platforms:                             2
  Platform Profile:                              FULL_PROFILE
  Platform Version:                              OpenCL 1.0 CUDA 3.2.1
  Platform Name:                                         NVIDIA CUDA
  Platform Vendor:                               NVIDIA Corporation


[msattler]

Just downloaded the new beta driver and installed it on my XP x64 dual 465 rig.....
Will see if it seems to do better than the 258.96.

The installer is pretty........but who gives a rip as long as it works.

Note to some......
I had noted earlier that the 258.96 XP 32 bit version would install correctly on my Win2K Advanced Server rig....
No longer......get a popup saying a newer version of Windows is required.
So hang onto your archived version of 258.96 if you have some older OSs in your farm.
It appears that nvidia has now officially abandoned support of them.

[Ghost0210]

Sneaky, nothing in the release notes or published about this going in that I've been able to find

Ahh, maybe they did announce the launch. Just re-read the Nvidia newsletter:

New Releases of Parallel Nsight and CUDA Toolkit

NVIDIA today announced new versions of the industry-leading developer tools Parallel Nsight and CUDA Toolkit.
– Parallel Nsight v1.5 includes support for Microsoft Visual Studio 2010, Tesla Compute
   Cluster (TCC) debugging and CUDA Toolkit 3.2
– CUDA Toolkit v3.2 includes new math libraries, performance improvements and support
   for the new 6GB Tesla and Quadro products
– For an overview of new features, see: http://is.gd/f9hKn

[Raistmer]

WoW! Great news indeed! In presentation they speak about FFT library performance improvement. If it's overall improvement (for all FFT types) it should directly increase CUDA MB
performance. (30% improvement in cuFFT if I heard it right)

ADDON: interesting info about Tesla's. If they would provide drivers to use secondary NV GPU in "tesla" mode (that is, separate driver, not graphics one) it would give great overhead reduce for GPUs too...