CUDA 4.1 + GeForce GTX 680 + Fedora 14
Hi all,

I've a Linux server with Fedora 14 64-bits and a GeForce GTX 680 (with 1536 CUDA cores) and other GPUs.

With the develop driver version 285.05.33, the system doesn't recognize the GTX 680.
For this, I have needed install the driver version 295.40. With that, the system finally recognizes the GTX 680.

But the problem is that the "deviceQuery" example of the SDK doesn't recognize the cuda cores of the GTX 680.
This is part of the output:

Device 3: "GeForce GTX 680"
CUDA Driver Version / Runtime Version 4.2 / 4.1
CUDA Capability Major/Minor version number: 3.0
Total amount of global memory: 2048 MBytes (2147287040 bytes)
[b]MapSMtoCores SM 3.0 is undefined (please update to the latest SDK)!
MapSMtoCores SM 3.0 is undefined (please update to the latest SDK)!
( 8) Multiprocessors x (-1) CUDA Cores/MP: -8 CUDA Cores[/b]
GPU Clock Speed: 0.71 GHz
Memory Clock rate: 3004.00 Mhz
Memory Bus Width: 256-bit


Is like if the Toolkit 4.1 or the SDK 4.1 don't support the CUDA capability 3.0.
Is possible that this be the problem?

Thanks a lot!!
Hi all,



I've a Linux server with Fedora 14 64-bits and a GeForce GTX 680 (with 1536 CUDA cores) and other GPUs.



With the develop driver version 285.05.33, the system doesn't recognize the GTX 680.

For this, I have needed install the driver version 295.40. With that, the system finally recognizes the GTX 680.



But the problem is that the "deviceQuery" example of the SDK doesn't recognize the cuda cores of the GTX 680.

This is part of the output:



Device 3: "GeForce GTX 680"

CUDA Driver Version / Runtime Version 4.2 / 4.1

CUDA Capability Major/Minor version number: 3.0

Total amount of global memory: 2048 MBytes (2147287040 bytes)

MapSMtoCores SM 3.0 is undefined (please update to the latest SDK)!

MapSMtoCores SM 3.0 is undefined (please update to the latest SDK)!

( 8) Multiprocessors x (-1) CUDA Cores/MP: -8 CUDA Cores


GPU Clock Speed: 0.71 GHz

Memory Clock rate: 3004.00 Mhz

Memory Bus Width: 256-bit





Is like if the Toolkit 4.1 or the SDK 4.1 don't support the CUDA capability 3.0.

Is possible that this be the problem?



Thanks a lot!!

GoParallel S.L. - http://www.goparallel.net

#1
Posted 04/13/2012 09:29 AM   
The CUDA API does not give a way for programs to ask how many CUDA cores per multiprocessor are present in a device. As a result, the deviceQuery example has to build in a lookup table to convert number of multiprocessors to number of CUDA cores. This has nothing to do with the toolkit or driver, so you can ignore the warning.

Note that a beta release of CUDA 4.2 was quietly announced on the forums:

http://forums.nvidia.com/index.php?showtopic=225731

You definitely want that toolkit, even though it won't update the SDK examples.
The CUDA API does not give a way for programs to ask how many CUDA cores per multiprocessor are present in a device. As a result, the deviceQuery example has to build in a lookup table to convert number of multiprocessors to number of CUDA cores. This has nothing to do with the toolkit or driver, so you can ignore the warning.



Note that a beta release of CUDA 4.2 was quietly announced on the forums:



http://forums.nvidia.com/index.php?showtopic=225731



You definitely want that toolkit, even though it won't update the SDK examples.

#2
Posted 04/13/2012 01:09 PM   
[quote name='seibert' date='13 April 2012 - 03:09 PM' timestamp='1334322565' post='1395788']
The CUDA API does not give a way for programs to ask how many CUDA cores per multiprocessor are present in a device. As a result, the deviceQuery example has to build in a lookup table to convert number of multiprocessors to number of CUDA cores. This has nothing to do with the toolkit or driver, so you can ignore the warning.

Note that a beta release of CUDA 4.2 was quietly announced on the forums:

http://forums.nvidia.com/index.php?showtopic=225731

You definitely want that toolkit, even though it won't update the SDK examples.
[/quote]


Effectively, I solved this problem with the installation the CUDA 4.2 (devdriver + Toolkit + SDK).
Now I only need to test if my CUDA program works fine on this GPU.

Thanks!!
[quote name='seibert' date='13 April 2012 - 03:09 PM' timestamp='1334322565' post='1395788']

The CUDA API does not give a way for programs to ask how many CUDA cores per multiprocessor are present in a device. As a result, the deviceQuery example has to build in a lookup table to convert number of multiprocessors to number of CUDA cores. This has nothing to do with the toolkit or driver, so you can ignore the warning.



Note that a beta release of CUDA 4.2 was quietly announced on the forums:



http://forums.nvidia.com/index.php?showtopic=225731



You definitely want that toolkit, even though it won't update the SDK examples.







Effectively, I solved this problem with the installation the CUDA 4.2 (devdriver + Toolkit + SDK).

Now I only need to test if my CUDA program works fine on this GPU.



Thanks!!

GoParallel S.L. - http://www.goparallel.net

#3
Posted 04/24/2012 09:56 AM   
I am interested to know which kind of GPGPU application it is, and how it's compare on GTX 680 vs. your old nVidia GPU (which model was it?) :p
I am interested to know which kind of GPGPU application it is, and how it's compare on GTX 680 vs. your old nVidia GPU (which model was it?) :p

Parallelis.com, Parallel-computing technologies and benchmarks. Current Projects: OpenCL Chess & OpenCL Benchmark

#4
Posted 04/24/2012 06:52 PM   
[quote name='parallelis' date='24 April 2012 - 08:52 PM' timestamp='1335293553' post='1400448']
I am interested to know which kind of GPGPU application it is, and how it's compare on GTX 680 vs. your old nVidia GPU (which model was it?) :p
[/quote]

Hi parallelis,

It is a micromagnetic simulator, but lamentably for now it not works properly with the GeForce GTX 680. We think it is because the last version of the CUDPP library that we used for some calculations of the program maybe is not compatible with the CUDA Capability 3.0.
Until now we have made satisfactory tests with Teslas 1060, 2050, 2070, 2075, 2090 and GeForce GTX 580.

You can get more information in our website: [url="http://www.goparallel.net"]http://www.goparallel.net[/url]
[quote name='parallelis' date='24 April 2012 - 08:52 PM' timestamp='1335293553' post='1400448']

I am interested to know which kind of GPGPU application it is, and how it's compare on GTX 680 vs. your old nVidia GPU (which model was it?) :p





Hi parallelis,



It is a micromagnetic simulator, but lamentably for now it not works properly with the GeForce GTX 680. We think it is because the last version of the CUDPP library that we used for some calculations of the program maybe is not compatible with the CUDA Capability 3.0.

Until now we have made satisfactory tests with Teslas 1060, 2050, 2070, 2075, 2090 and GeForce GTX 580.



You can get more information in our website: http://www.goparallel.net

GoParallel S.L. - http://www.goparallel.net

#5
Posted 04/25/2012 08:16 AM   
Exactly, is needed install the CUDPP 2.0 to run CUDA programs with the CUDA Toolkit 4.2 and CUDA Capability 3.0

But now, the problem is that the program runs more slowly with the new driver than with the old driver.
With the driver 270.41.19 is much faster than with the driver 285.05.33, 295.40, 295.41 and 295.49.

Is a driver problem? Any idea please?
Exactly, is needed install the CUDPP 2.0 to run CUDA programs with the CUDA Toolkit 4.2 and CUDA Capability 3.0



But now, the problem is that the program runs more slowly with the new driver than with the old driver.

With the driver 270.41.19 is much faster than with the driver 285.05.33, 295.40, 295.41 and 295.49.



Is a driver problem? Any idea please?

GoParallel S.L. - http://www.goparallel.net

#6
Posted 05/11/2012 09:27 AM   
Scroll To Top