GPU Perfomance How much GFlops???

[Dmitry_Russky]

…

Mac, i thik that is something wrong on my enviroment. Can you plese run the next test on your PC.

simply replace the files in “simpleMultiGPU” SDK example and execute.

I am realy confused. Where am I wrong?

[Dmitry_Russky]

One more file.
CPP have to be changed to .cu

[Dmitry_Russky]

Dmitry,

I’ve looked at AMD’s APIs and found those harder to use, than NVidia’s CUDA. This is the primary reason why I’m using CUDA and buying NVidia’s hardware, even though I’ve heard ATI’s boards are 20% or so faster (I can offer no evidence to this last statement). I am under the impression, that I’m in the majority (NOT just on this website, but among GPGPU developers in general) regarding the ease of use assessment.

I understand and agree with you. But it depends on application. If finaly i will get 800% (8 times) better perfomance, i am ready to spend time to understand Stream. Again, it depend in which field you are working.

For example development of ASIC takes years, but finaly aniway it wins. :-)

As far as my application is concerned, memory bandwidth is not an issue, but efficient AND EASY TO CODE access to the registers and shared memory is: contrary to your FTP example, my code performs very many integer and floating point operations per global memory read/write. I don’t find brook to be a convenient tool for writing complex massively multi-threaded code against the shared memory.

I think that is a missunderstaning in concept.

IHMO, GPU it is like FPGA with one difference. If in FPGA everithing have to be done in VHDL, in GPU simple parts have to be done in CUDA or Streams(Small logic cells, shaders, could be programed in C language), and complicated calculation in high lavel language. For me GPU is like a co Processor. Execution have to be done in GPU, managing have to be done in CPU. For that should be possible to execute kernels very fast. To save memory bandwidth, intermidiet results between calls of different functions, could be stored in shared memory, and so on… But, anyway, i need my GFlops! :-)

I would be very interested in seeing a piece of complex numerical code (e.g. shared memory-based solver of a VERY LARGE number of different SMALL linear systems of equations, each of which fully fits into the shared memory), that works much faster on an ATI, than on NVidia hardware of the same generation. I would be particularly impressed, if this ATI code appears to be easier to implement, than the CUDA code.

I promise, you will see. :-)

Just give me a time . First, i will implement 1D FFT and FIR filter. I like ATI Streams more because I need more then just APIs. :-)

[_Big_Mac]

Dimitri, the way you launch your kernels (going through 1 - 10 blocks and 1 - 256 threads) is really bad. First of, 10 blocks won’t even saturate your GPU. You’ll want hundreds if not thousands. Secondly, one should launch a multiple of 32 threads per block. Otherwise you’re loosing performance because the blocks are not aligned to warp size (SIMD vector length).

Lastly, unrolling the loop 1024 times instead of ex. 200 will likely hurt performance because the executable will get too big to fit into the GPU’s instruction cache.

Tomorrow I will write a program that runs the kernels as they’re supposed to and writes flops to console. I will give you the whole project and a binary just in case. Are you using Visual Studio 2008?

[Dmitry_Russky]

Tomorrow I will write a program that runs the kernels as they’re supposed to and writes flops to console. I will give you the whole project and a binary just in case. Are you using Visual Studio 2008?

Super!

Thank you very much!

Yes, i am using MSVS2005 and 2008.

If i will be able get the perfomance, will be very good. :rolleyes:

Thanks,

Dmitry

[_Big_Mac]

Here’s the project.
[attachment=10314:flopsBenchmark.rar]

Requires:

• Visual Studio 2008 and Cuda.Rules file.
  An easy way to add the rules file is by installing [url=“CUDA VS Wizard download | SourceForge.net”]http://sourceforge.net/projects/cudavswizard/[/url]
• CUDA 2.3

[_Big_Mac]

GTX 260 results (theoretical peak/obtained)
mad 545/524, 96% efficiency
mad+mul 817/576, 70% efficiency

My 8800 GTS results
mad 414/403, 97% efficiency
mad+mul 622/458, 74% efficiency

[Dmitry_Russky]

Here’s the project.

[attachment=13320:flopsBenchmark.rar]

Requires:

• Visual Studio 2008 and Cuda.Rules file.

An easy way to add the rules file is by installing http://sourceforge.net/projects/cudavswizard/

• CUDA 2.3

Hi Big Mac!

Finaly i have made a tests, and i rely got my GFlops. (476/445, and 458/715).

Thank you very much!

I will made some tests soon, and will try to find the reason why it was not possible to get flops before.

Again, thanks a lot!

Dmitry