Siegmar Szlavik wrote:

speaking of FLOPS, besides my ~3 MHz P4, which can perform ~6 GFLOPS, I have
a much more powerfull processor in my system. It's the NV40 based GPU on the
graphics card and due to its 16 pixel processing piplines the compared FLOP
performance would be that of a >10 GHz P4 CPU. The GPGPU website
(http://www.gpgpu.org) deals with the subject of general purpose
computations on the GPU and especially the scientific computing fraction is
well represented. Did anyone ever thought of using the GPU-Power for
LL-testing?

I'd imagine the biggest issue would be precision. But, this paper does look intriguing.... http://www.cs.unm.edu/~kmorel/documents/fftgpu/fftgpu.pdf


They manage to achieve a 0.37Hz (about 3 seconds per iteration) 2D FFT of 1024x1024 4-channel 32bit floating data, for a computation rate of about 2.5GFLOPS.. not bad. catch-22, the NVIDIA graphics they use in their evaluation is a $400+ video card.


_______________________________________________ Prime mailing list [EMAIL PROTECTED] http://hogranch.com/mailman/listinfo/prime

Reply via email to