Running AMD card with CUDAFY.net

Jul 5, 2013 at 8:09 AM
Hi, from what i have understood, its possible to run Cudafy.net on AMD cards just by selecting openCL mode? Its also appears that the AMD cards seem to have a vast performance improvement compared to the GTX gaming cards Benchmark.

Currently im running a Gtx 660 sc which im using for general computing. Unfortunately i cant justify spending several thousand pounds on the likes of a quadro, tesla or firepro, so i was wondering if annyone is currently using an AMD card with Cudafy.net, are the figures in the above link actually true, and is there something im missing as it seems the AMD HD7970 is a no brainer at £470.
Coordinator
Jul 5, 2013 at 2:37 PM
I've used an entry level AMD card with CUDAfy.NET, not tested with higher end cards. Consensus appears to be "it depends". Depends on algorithm, single or double, etc. CUDAfy has little overhead above CUDA or OpenCL, so best read articles on compute performance of similar algorithms to yours.
Jul 5, 2013 at 4:42 PM
So its not just a matter of comparing single (not using double atm, althought this may change) TFlops performance? From what i had read, some people say that Nvidia have tailored their gaming cards towards single precision performance. I found the following online which seems to suggest they have:


AMD 7970 Single = 3788 GFlops Double = 947 GFlops

GTX 770 Single = 3213 GFlops Double = 134 GFlops

Im assuming since you havent used a high end AMD card, that your personal preference is for Nvidia cards?

Thanks for your help :)
Coordinator
Jul 5, 2013 at 7:42 PM
Edited Jul 5, 2013 at 7:46 PM
My preference for NVIDIA is down to CUDA which is much more refined and easy to use than OpenCL. Even in CUDAfy which takes care of most of the OpenCL nightmare (think CUDA driver API on a bad day) there are limitations compared to targeting CUDA devices: e.g. functions in structs, vendor specific stuff, no maths libraries (FFT, BLAS, RAND etc), etc.
There is much more to it than Tflops, but in general if you use double you're better off with AMD.
Here is an useful comparison over at AnandTech.
Jul 6, 2013 at 9:44 AM
Thanks for the heads up regarding CUDA vs OpenCL. Currently im not using any of the functionality you mentioned but in the off chance i do later, i guess i will be kicking myself.