questions about amd apu fx opencl

[ev56o]

i' m assembling an high performance asus amd nvidia pc for gaming and i don' t know what processor choose.
is the apu a solution dedicated to high performance (with very big pcie 32x 3.0 videocard) or for economic pc?
are linux, opencl and the applications compatible and well performed with any processor amd apu fx?
in the sense it will use video card and apu together for improve performance?

[keet]

I'd look up both C.P.U. and graphics benchmarks for the particular model of A.P.U. that you are considering; that's probably the best way to get a clear answer. In general, though, I'd guess that the A.P.U.s don't perform quite as well as discrete graphics cards, but they're probably enough for basic to moderate 3D rendering. Integrated graphics are basically never as good as a medium-end discrete graphics card. I am not sure whether any setup would use the A.P.U. and the graphics card simultaneously to build on each others' performance, but I doubt it; it doesn't seem like something that Crossfire would support.

[ev56o]

yes i confirm i read on web that the apu generation is going end and was made for economic solution.
i add a new doubt from the site asus for a88x-plus fm2+

[vaxbrat]

The latest AMD APU's at the moment, Kaveri, are interesting beasts. They are Steamroller core based for the processor so have a couple of extra new processor features compared with the older Piledriver generation cores being used in the FX-8xxx and FX-9xxx processor boards. However there are only 4 cores versus the 8 in the FX boards.

I think Crossfire will blend a discrete AMD display card with the on-board graphics, but you will want to keep the arch's as close as possible. Kaveri is basically a Hawaii class embedded graphics adapter so you will want something GCN based such as a Bonaire, Pitcairne, Tahiiti or Hawaii based card. In terms of model numbers, you are probably looking at only AMD cards that have been released in the last 1-2 years for doing Crossfire with Kaveri.

If you use Nvidia, you will end up disabling the on-board gpu since they do SLI instead of Crossfire.

The real holy grail about AMD APU's starting with Kaveri and the upcoming release is the HSA architecture. GPU programmers have always had to deal with the headache of marshalling data between the CPU and the GPU. With HSA, those problems all go away. These may end up being very popular in the supercomputer community for those folk who do massive gpu processing oriented clusters.

[ct85711]

I also have an APU on my system, and I found a benchmark that was supposedly done with apu systems with and without a discrete video card and a intel system with a amd video card. In short the benchmark displayed a interesting part, in that overall the APU processors are not mature enough to have a advantage. Simply put, it showed running a APU with a compatible video card run worse together than working together. From what the results showed, most games that typically have a strong crossfire support showed worse performance with the APU working with a discrete working (one that is supposedly supported). Of course the benchmark was done on a windows system (where you'd expect the best driver support available). So overall, the APU is just like a cpu with a dissected video card combined together without the memory bandwidth to do a true video card's work.

[szatox]

Actually crossfire's performance depends heavily on the mode you run it. I saw for example benchmark where each device in APU + GPU tandem was drawing every second frame. And this mode was almost as good as double performance of the weaker device. Drawing 2 halves of a single frame by 2 devices hardly gave any performance gain though.

Anyway, for gaming dedicated GPU will always be better than APU, because gaming is not the purpose of APU. Those things had been designed to be cheap and be enough for most cases, and they perform very well for their price. You can watch a film using it. You can even play some games. Performance will be worse, but might still be acceptable.
Relative performance in case of OpenCL computing seems to deppend on how much data you transfer between host's RAM and GPU. RAM bandwidth is higher than PCIe, but dedicated RAM will of course do better between transfers. Is the program you run on GPU short enough (And using enough data) to outperform I/O? Or maybe you simply need more shaders? I guess second case will be much more common. Still, wjen I tried bitcoin mining on my APU, CPU part was 4x2Mhps and GPU part was 80Mhps.