I worked for a time with Ati Stream on my 4850, but now I think its time to switch to OpenCl.
Because I want to build up a new (2nd) computer, i am interested in the OpenCl performance difference of a 4850 and a 5750, because performance in 3d games is nearly the same (4850 is cheeper) , but I am not shure about, how much more the 5750 is optimized for OpenCl use.
So I want to ask for some benchmarks - if any exist
The technique of local memory will be faster on a 5xxx series, since it is emulated on a 4xxx. If the kernels you write/run do not use that technique then you might not notice much. 4xxx is more legacy than built to run OpenCL, so there are probably other things too.
If speed per $ is your primary concern, I am not confident I can answer that.
I have compared the performance of beta4-SDK samples between a 4870 and 5770 and found that 5770 was faster in most of the samples.
The performance difference may be a bit larger in case a sample is using local memory. This is because on R7xx series this leads to an extra copy between global<->global as local memory is emulated by global buffer, this will infact deteriorate the performance further compared to R8xx which has the 32KB LDS available in opencl.
The ALU performance is nearly same between the two cards in OpenCL. I can give you specific benchmark results if you want like : ALU performance, Global memory read/write performance, local memory performance etc.