Daniel O'Connor wrote: > On Friday 15 June 2007 04:52, Rohit Garg wrote: > >>Has any body tried porting the compute intensive portions to GPU's. >>The task apparantly is ideal for GPU's to digest. I know that GPGPU >>is, as yet, more research and less apps area, but I feel that it will >>be valuable addition. Yes I have done so a whila ago.
I ported some FIR filters to the GPU. The results were somewhat promising but not very good. I got about 500 Mtaps/sec on a nvidia 5700LE after tuning the memory access patterns. memory bandwidth and AGP/PCI-E bandwidth are the main issues. Another issue is that gnuradio data is 1-dimensional. GPU's do very well in 2-dimensional data but you have to jump a lot of loopholes to get 1-dimensional data to work en perform well. (The cache in a GPU is also 2-dimensional) With nvidia CUDA and/or ATI CTM and recent hardware results should probably be much better for FIR filters and things like that. Blocks that work with a loopback (like IIR, PLLs) will never work well on a GPU because you cannot parralellize the algorithm. Greetings, Martin > > > http://www.google.com/search?q=gnuradio+gpu&ie=UTF-8&oe=UTF-8 > > :) > > > > ------------------------------------------------------------------------ > > _______________________________________________ > Discuss-gnuradio mailing list > [email protected] > http://lists.gnu.org/mailman/listinfo/discuss-gnuradio _______________________________________________ Discuss-gnuradio mailing list [email protected] http://lists.gnu.org/mailman/listinfo/discuss-gnuradio
