Re: I am kinda impressed...
Hey Beer,
I ran some Collatz recently.
Sampling: 9 collatz v2.09 (ati13ati) wu's (That's all that was still listed)
Avg. GPU Runtime: 1,659.07 sec.
Credit: 3,131.88 per wu
GPU: Radeon HD 5830 (1GB DDR5) - Core OC: 925Mhz
Re: I am kinda impressed...
Collatz does much better on AMD than nVidia GPUs.
1. It is partly due to the simplicity of the Collatz algorithm. nVidia GPUs tend to have fewer but more powerful (larger instruction set) stream processors whereas AMD tends to have more but simpler stream processors.
2. It also has to do with how much data a project needs to copy to/from the GPU. nVidia beats AMD in that category, but the Collatz app relies more on reading from texture memory than copying data to/from the system RAM and AMD has faster texture mapping.
3. Lastly, it also has to do with Gipsel having tweaked and tuned the IL code (aka GPU assembly language) for the ATI app to make it 30% faster than the AMD Brook+ compiler's output. If there is a PTX guru out that that wants to volunteer to tweak the CUDA code, I'd be happy to work with them. But, due to #1 and #2, the CUDA apps probably won't ever match the speed of the ATI apps - or at least not until AMD forces everyone to use OpenCL.
It also goes to show that cross project parity can't ever happen because different projects work better on different hardware. Since PG works better on CUDA, would it be fair to grant the AMD GPUs less credit on Collatz so that it matched PG's AMD credit? Or, since Donate works so much better on AMD, should PG raise their AMD credit even though the nVidia GPUs work better?