5850/5870/5970 Folding@Home Support - When?

[exar333]

Anyone know when a new version of F@H will be released supporting the 5xxx series? I am currently doing the "workaround" that tricks the program into thinking your GPU is actually a 4870, but the folding performance isn't much to write home about. It is substantially slower than my old GTX285.

Don't get me wrong, I love the 5870, but I was somewhat suprised about the poor F@H support it has, especially since it has been out for a couple months.

 

[Creig]

our focus is now on the next generation GPU3 client and OpenMM core. The programming language and model used by the old ATI code have been made largely obsolete by the introduction of OpenCL, so we're focusing our efforts there, rather than trying to revamp an older code.

http://foldingforum.org/viewtopic.php?f=51&t=12049&start=15


I would imagine that 5000 series support will be included in the new GPU3 client. As to when exactly it will be released, I heard sometime Q1 of this year.

 

[Denithor]

So does this mean ATI performance will finally "catch up" to nVidia (similar game performance hardware = similar ppd output)? Or will the basic hardware differences continue to allow nVidia to hold a substantial lead?

 

[dook43]

So does this mean ATI performance will finally "catch up" to nVidia (similar game performance hardware = similar ppd output)? Or will the basic hardware differences continue to allow nVidia to hold a substantial lead?

Supposedly the 4800 series cards were far ahead of GT200 in theoretical performance but Stream was holding them back.

I'd look for a properly optimized OpenCL F@H build allowing ATI to wax the floor with Nvidia.

 

[nismotigerwvu]

Supposedly the 4800 series cards were far ahead of GT200 in theoretical performance but Stream was holding them back.

I'd look for a properly optimized OpenCL F@H build allowing ATI to wax the floor with Nvidia.

I'm not sure sure about waxing them, as we've already seen some reports that openCL code on 4000 series hardware requires some specific parameters to actually run at reasonable speeds. If it is written in a generic form, these cards may suffer. As far as the 5000 cards go I can't quite remember, but I'm sure someone more informed will chime in after me.

 

[exar333]

It would be great to see AMD put some resources into optimizing the F@H client to really fly on the 5xxx series. It could be some good publicity, and might actually sway a few buyers their way while enabling a good cause.

 

[Creig]

So does this mean ATI performance will finally "catch up" to nVidia (similar game performance hardware = similar ppd output)? Or will the basic hardware differences continue to allow nVidia to hold a substantial lead?

I believe we will see a two-fold reason for an ATi performance increase in the furue:

1. The ATI client was developed before either CAL or CUDA existed. It was intially designed to use Brook as an interface with DirectX and then further adapted as GPU2 to avoid the problems assoicated with DirectX. The NVidia core was developed without the Brook interface which avoids a certain amount of extra overhead.

2. I believe that WU size also currently favors Nvidia architecture, but that the size is slowly decreasing which will favor ATi cards more and more as the project progresses.

Please don't take #2 as pure fact as that was recalled simply from memory. I couldn't find a link to back it up, but believe it to be a true statement.

 

[nyker96]

I bet they will use the new OpenCL api for the version for hd5xxx series.

 

[GodisanAtheist]

ATI recently admitted that the fundamental design of the 4xxx series cards simply does not allow it to perform well in GPGPU tasks. If I'm not mistaken, the issue is the lack of a serious local cache for each of the SIMD clusters, so API's like OpenCL have to cordon off space in the global cache and "pretend" like each of these spaces is actually local cache for each of the SIMD clusters. This is actually a workable solution so long as the crap you're caching is smaller than the allocated space, if not it sends performance into the dumpster.

The 5xxx series cards got significantly beefier local caches and tend to bulldoze OpenCL *BENCHMARKS*. We'll see how the real thing goes.

I do agree that ATI should really get on the ball here. While I've never been a big fan of F@H, I'm sure a significant minority of enthusiasts base their purchasing decisions on proper support for F@H.

 

[exar333]

ATI recently admitted that the fundamental design of the 4xxx series cards simply does not allow it to perform well in GPGPU tasks. If I'm not mistaken, the issue is the lack of a serious local cache for each of the SIMD clusters, so API's like OpenCL have to cordon off space in the global cache and "pretend" like each of these spaces is actually local cache for each of the SIMD clusters. This is actually a workable solution so long as the crap you're caching is smaller than the allocated space, if not it sends performance into the dumpster.

The 5xxx series cards got significantly beefier local caches and tend to bulldoze OpenCL *BENCHMARKS*. We'll see how the real thing goes.

I do agree that ATI should really get on the ball here. While I've never been a big fan of F@H, I'm sure a significant minority of enthusiasts base their purchasing decisions on proper support for F@H.

Great explanation, thanks!