[bchur83]

I recently started with Folding, and have a question about the GPU client. I have been having trouble with running the SMP CPU client with the GPU client. The GPU client uses 50% of the CPU usage and the other 50% is shared between the clients on the SMP side. Is this normal? I have an E6600 running at 3.35Ghz, and seem to be too slow on the SMP side to finish units before the due date. Should I drop the SMP client for a standard client? I have also been doing some video converting so the CPU has been loaded most of the time with that. I set the GPU to below normal priority while converting videos so it continues to run.

[turbohans]

uhhh, mine usually uses between 1 and 5% cpu!! not shure whats going on there!?!?

are you shure thats your gpu core your looking at?? normally its fahcore_14.exe under processes in your TM or RM.

[cmichaelt]

What is the OS you are currently running?

[Evilsizer]

if your running xp with ati you need to allow one core for the gpu. then just run a single client console for the other core. now if this was NV based gpu then from what i have read there is no more cpu usage. from the past when ?nitteo? did some testing he said he could get lower cpu usage on Vista then with XP. im hoping that in future drivers that cpu usage will be 0 like with NV when doing GPU client with ATI cards.

[bchur83]

OK, that explains it. I am running a 4870 on XP so I will tell the SMP to just use a single client.

Thanks guys.

[harlam357]

You may be able to use an app like SetAffinityII and hold the SMP core processes on core 0 of your cpu and let the GPU2 core use core 1 of your cpu.

I used to do this on my E8400 @ 4.05GHz before the NV drivers got better and required little to no cpu usage... I made the SMP deadlines no sweat. I imaging that E6600 would do similarly well... just be sure it makes the preferred deadline.

[Bon3thugz43v3r]

If you want to get your cpu usage down basically you need to add environmental variables.

See this thread

With a 4870 I am going to suggest these settings.

CAL_NO_FLUSH 1
BROOK_YIELD 2
FLUSH_INTERVAL 128

[Arkaine23]

Bon3thugz43v3r: if only that actually worked. A lot of people are still seeing 100% cpu utilization with those variables set.... and tweaking flush_interval higher and lower as well.

Mine dropped cpu useage to 5% when I first installed the newer driver and got the newer fah_core. But ever since that first WU's its been back to 100% cpu no matter what I try. Others are reporting similar issues.

[Evilsizer]

i gave up on the ATI/GPU thing, to many VPU recoveries going on in XP-sp3. i got another box to setup with win7 beta. hopefully i dont get constant VPU recoveries there...

[bchur83]

Yeah, I have been getting a lot of VPU recoveries since running the GPU client. At least 2-4 a day during normal use. Currently I am just running the GPU client on one core and a single instance of the standard client on the other. I have been eying getting a quad core and then not having to worry about it. We will see.

[Jim Morbid]

Whats at the root of this? Is it a lack of CUDA or is the Forceware drive just far superior for Folding?

[Bon3thugz43v3r]

Originally Posted by Arkaine23

Bon3thugz43v3r: if only that actually worked. A lot of people are still seeing 100% cpu utilization with those variables set.... and tweaking flush_interval higher and lower as well.

Mine dropped cpu useage to 5% when I first installed the newer driver and got the newer fah_core. But ever since that first WU's its been back to 100% cpu no matter what I try. Others are reporting similar issues.

Arkaine23, I never got 100% CPU utilization without the variables. The CPU utilization would be around 40% to 50%. What O/S do you use?

When adding the variables I have dropped FAHCORE_11 to only utilize the cpu when it is needed, not the whole time. Example: My usage on project 5739 is under 10%. I have a 3.6ghz e8400.

CAL_NO_FLUSH 1 BROOK_YIELD 2 FLUSH_INTERVAL 128

Now I want to add this because I just set up Notfred's VM client yesterday.
The virtual machine defaults to normal in task manger for priority. Because FAHCORE_11 is defaulted at low this means that VM's will STOP GPU folding unless you manually change it every time. I never had this issue when running MPICH x64 with -SMP flag obviously since the FahCore_A1's would default at low in my task manager.

I am pulling 2,700 PPD with Notfreds VMs and 2,700 PPD GPU on the worse ppd project handed out by stanford for the GPU2 client (Project 5739). With projects 475x (477 pointers) My GPU is putting out 3,900 PPD and VMs 2,700 PPD.

My point in mentioning all this is that if you take the time to configure the variables for what works for your system like I have, then you will see huge ppd increases which in turn will help team 32 remain in the top5.

I just couldn't settle for less.

[Bon3thugz43v3r]

Originally Posted by Jim Morbid

Whats at the root of this? Is it a lack of CUDA or is the Forceware drive just far superior for Folding?

This is from stanfords website.

ATI (2005-present)

We've been working with ATI for quite a while on our GPU core. This started with GPU1 and has carried over to GPU2.

NVIDIA (2007-present)

We've been working with NVIDIA on the GPU2 port for NVIDIA hardware. In a collaboration between NVIDIA personnel (Scott LeGrand) the Folding@home team, we have ported and optimized our code to CUDA.

Now, here is the thing. Since 2007 ATI and Stanford have done very little in optimizing the original stuff from back in 05. I read on the forums that they were close to starting single card optimizations for ATI. Such as the 4800 series. Who knows when this will happen though.

CUDA is the reason why Nvidia cards are cranking out 768 pointers in 4 hours.

[Arkaine23]

I'm using XP SP3. Cat 9.4. Its a 2600pro AGP card coupled with a 1mb cache single core 2.2Ghz Turion CPU.

The GPU PPD is about the same as its always been... just short of 500 PPD. I was hoping to get a massive +200 PPD by being able to run a uniprocessor client along with the GPU client.

[Bon3thugz43v3r]

Ah, that explains everything. Single core 1mb cache