An evaluation configuration Linux, FFT, hardware

Hello,
I wander if someone can help me to setup an hardware configuration for evaluation of the GPU computing in my problem.
My need is to make plenty of Raw data as 2x8192 reals to 8192 complex FFT. By plenty I mean that currently I get 10kHz data that I try to save on disk and make the FFT afterwards. My field of interest is radioastronomy time signal.
So my questions are

  1. is a GPU FFT (as CUDA provided) can help to perform the FFT in paralell ?
  2. which minimal hardware should I take to test : GeForce GTX 480 ? or Telsa, which motherboard? do I need a graphic baoard for display and an other one for computing ?
  3. I am using Linux Ubuntu 10.04, is-it a suatable linux distrib. is this context?

Thanks in advance for any comments and advises.
Jean-Eric

Hello,
I wander if someone can help me to setup an hardware configuration for evaluation of the GPU computing in my problem.
My need is to make plenty of Raw data as 2x8192 reals to 8192 complex FFT. By plenty I mean that currently I get 10kHz data that I try to save on disk and make the FFT afterwards. My field of interest is radioastronomy time signal.
So my questions are

  1. is a GPU FFT (as CUDA provided) can help to perform the FFT in paralell ?
  2. which minimal hardware should I take to test : GeForce GTX 480 ? or Telsa, which motherboard? do I need a graphic baoard for display and an other one for computing ?
  3. I am using Linux Ubuntu 10.04, is-it a suatable linux distrib. is this context?

Thanks in advance for any comments and advises.
Jean-Eric

Yes, NVIDIA provides a CUFFT library to perform FFTs on NVIDIA GPUs. Given the relatively small size of your FFTs, you will want to load many waveforms and perform a batch 1D FFT to efficiently use the full capability of the GPU. Do you have any processing latency requirements that could make batching a problem?

If you are just starting out, I would suggest the GTX 480 or GTX 470. (The electrical power requirements of the GTX 470 are lower, so it might be easier to install.) I don’t think you need a Tesla card for testing, as it will not run any faster. You can use the same card for display and computing if your individual CUDA kernels run less than 5 seconds, or if you turn off X11. I generally run CUDA programs over SSH on a computer with no display.

Yes. Ubuntu 10.04 is supported by CUDA 3.2, which is currently at the release candidate 2 stage and will hopefully have a final release soon.

Yes, NVIDIA provides a CUFFT library to perform FFTs on NVIDIA GPUs. Given the relatively small size of your FFTs, you will want to load many waveforms and perform a batch 1D FFT to efficiently use the full capability of the GPU. Do you have any processing latency requirements that could make batching a problem?

If you are just starting out, I would suggest the GTX 480 or GTX 470. (The electrical power requirements of the GTX 470 are lower, so it might be easier to install.) I don’t think you need a Tesla card for testing, as it will not run any faster. You can use the same card for display and computing if your individual CUDA kernels run less than 5 seconds, or if you turn off X11. I generally run CUDA programs over SSH on a computer with no display.

Yes. Ubuntu 10.04 is supported by CUDA 3.2, which is currently at the release candidate 2 stage and will hopefully have a final release soon.

I can load for instance 5120 samples and batch them ? is it correct?

I have a GTX 460 does it still fit ?

How do you proceed to “run CUDA programs over ssh on a computer wo display”, do you meen that you connect to the computer (wo display) via ssh through network?

I have seen some “Howto” on the web to install CUDA on Ubuntu, but can you give me a safe procedure that you follow (it may be a web page)?

Thanks in advance for your help.

JE

I can load for instance 5120 samples and batch them ? is it correct?

I have a GTX 460 does it still fit ?

How do you proceed to “run CUDA programs over ssh on a computer wo display”, do you meen that you connect to the computer (wo display) via ssh through network?

I have seen some “Howto” on the web to install CUDA on Ubuntu, but can you give me a safe procedure that you follow (it may be a web page)?

Thanks in advance for your help.

JE

You should read through the CUFFT Users Guide to make sure:

http://developer.download.nvidia.com/compute/cuda/3_2/toolkit/docs/CUFFT_Library.pdf

(I’ve only used PyCUDA + PyFFT for a small project personally, so I don’t know CUFFT’s limits.)

Yes, a GTX 460 will also work fine.

Correct. The GPU workstation runs an SSH server and I connect to that and run my programs on the command line.

I don’t really have a special procedure. I go to here and download the required packages:

http://developer.nvidia.com/object/cuda_3_2_toolkit_rc.html

NVIDIA has a Getting Started Guide for Linux:

http://developer.download.nvidia.com/compute/cuda/3_2/docs/Getting_Started_Linux.pdf

which is probably a good thing to read and follow.

You should read through the CUFFT Users Guide to make sure:

http://developer.download.nvidia.com/compute/cuda/3_2/toolkit/docs/CUFFT_Library.pdf

(I’ve only used PyCUDA + PyFFT for a small project personally, so I don’t know CUFFT’s limits.)

Yes, a GTX 460 will also work fine.

Correct. The GPU workstation runs an SSH server and I connect to that and run my programs on the command line.

I don’t really have a special procedure. I go to here and download the required packages:

http://developer.nvidia.com/object/cuda_3_2_toolkit_rc.html

NVIDIA has a Getting Started Guide for Linux:

http://developer.download.nvidia.com/compute/cuda/3_2/docs/Getting_Started_Linux.pdf

which is probably a good thing to read and follow.

Thanks. I will come back as soon as I new results.

Thanks. I will come back as soon as I new results.

On CUFFT documentation limits for FFT are :

1D FFT -> up to 8,000,000 elements
2D and 3D -> [ 2 16384 ]

On CUFFT documentation limits for FFT are :

1D FFT -> up to 8,000,000 elements
2D and 3D -> [ 2 16384 ]