Reducing CPU Usage from AMD GPU Folding

[Smith6612]

Heya guys,

 

I've been folding for quite some time, although I've not started contributing to the Linus Tech Tips group yet. I wanted some help on getting GPU folding to run more efficiently on some systems here. Most of my systems here are on NVIDIA GPUs, and as a result they do not suffer from the infamous 1 core 100% usage issue that the AMD GPUs seem to suffer. Those NVIDIA GPUs when folding use maybe 1% of the CPU, meaning I can pretty much run an SMP client to all cores.

 

Anyways, I have a few systems at home that I believe could have more folding potential. One of these systems has an AMD APU in it which is also being used for folding. Specifically, the AMD A6-4320M with Radeon HD 6520g Graphics. The machine for the most part works fine, however when running SMP units it suffers from 100% CPU usage on a single core which limits me to 3 out of the 4 presented cores on this CPU (Yeah, I know these things are 1.5 cores due to the CPU being based around Bulldozer). Anyways, it gets around 2,500PPD folding on SMP alone, and GPU Folding can get around 1,500PPD with it running alone on the non-optimized 0x16 core (0x17 gets a bit more, but it's in beta at this time and this system does not fold 24/7). I'm looking to get around 3,000PPD on this chip at minimum.

 

After searching online a bit, it seems that there are some environmental variables you can set for AMD Cards to reduce CPU usage, however most of what I've found is for older ATI stream compute cores, and not the newer OpenCL-based cores. So essentially what my question is, would these environmental variables still apply, even if the AMD APP SDK isn't installed, and if so, which variables should be used?

 

Thanks.

[Whaler_99]

Can you provide an example of said variables for us to look at? I used to play with some software that tweaked AMD cards for Xfire performance years ago, but nothing recently and certainly not for folding.

[warmon6]

Man, I've totally forgot about that.... Then again, I dont use AMD/ATI cards so I guess that was going to happen...

 

 

Here's the variables smith is talking about.

 

http://en.fah-addict.net/articles/articles-1-3+gpu-environment-variables.php

 

http://fahwiki.net/index.php/Environment_Variables_for_the_GPU_Client

 

 

Although I have no idea if the new core is affect by this method like the older core were or if it is, nor do i know if it would be better or worse but feel free to try! 

[Whaler_99]

Ahhh... interesting... well, they both reference some pretty old versions - I think this may have been centered around getting your GPU to run at max performance. Not sure how relevant they are with the newest drivers and the new client. When the "slider" is at full, most cards are very good at ramping up to 100%

 

On my gaming system, with the older 7.x client, even with gpu at "full" load, I could still use the computer fine and even play some light games. With the latest 7.3.x client, this is no longer the case. When running GPU at full, I can noticed some delays doing stuff like watching Youtube and gaming is impossible unless I pause the GPU fold. This tells me they have definitely tweaked the GPU component of folding.

 

You could try testing it though - but remember, only change one variable and run that for a few days to get a good baseline and hopefully how different projects are effected.

[Smith6612]

Thanks. For any of the dedicated folders out here, if we could definitely give some of the variables a try (those were the ones by the way, warmon6!), I'd love to see what sort of results we get especially on machines running multiple 7970s and so on. For now, I'm letting the APU fold until this weekend, since I've read some posts about the Environmental variables causing some BSODs. Given the machine is used by many, I'd hate for a BSOD to just get up and happen. One support call for me.

 

It definitely seems like Brook_yield would be the one I'd want to try to use the most here combined with Flush_Interval, although despite 100% load on the IGP the system UI remains responsive as is with the default settings, and is fast enough still to run games on. Considering the IGP isn't nearly as fast as a dedicated GPU, I don't see the need to tie up the CPU waiting for the IGP to finish processing a batch of data.

[warmon6]

Glad to be of some assistance even with my lack of knowledge of AMD cards when it come's to F@H. 

[Smith6612]

So, I gave this a try. It seems these settings still work with Core 15, however as mentioned, they only adjusted GPU usage. It seems the problem is with the drivers when OpenCL is being called at this point (should have known). Although, I know NVIDIA has been suffering the same problems with their OpenCL drivers, too when you run Core 17 for GPUs. In order to keep the GPU within this APU folding, I had to set the SMP client back to 3 out of 4 cores as otherwise, the GPU core would be starved.

 

Hoping they fix this. Come Core 17, SMP clients are going to have to be backed off a core as the Pende team won't have CUDA support in that core.