To the coders of P7149,7156 - well done!

[Simba123]

One thing that has annoyed me greatly in Folding is that adding a GPU client can
trash the ppd of the SMP client by 20-30% .

I have noticed that P7149, P7156 barely drop 10% when a GPU client is added when they are
folding. This is awesome.

I don't know why they fold so well with GPU, but they do. This should be investigated and encouraged
because it greatly increases the efficiency of folding both GPU and SMP at the same time.

rig is 17 2600K@4.5 SMP8 using V6(Tracker) client, 560ti 2Gb. Win7 SP1, nVidia 301.42

Example P7149 on SMP8 produces 31265ppd @ 2:19 tpf
Add P8009 GPU (no change to smp8)
the GPU produces 17489
the smp drops to 28486.

The same gpu project added when a P6940 is running results in PPD going from 29042 down to 23900.

[Jesse_V]

Hmm. My 560TI uses at most 5% of the CPU (also a 2600K) while folding. ATI GPUs use one whole CPU core, but Nvidia GPUs use much less.

[bruce]

All ATI gpus cause smp to slow down due to the loss of one CPU. NVIDIA gpus do not. It's not the project.

[derrickmcc]

rig is 17 2600K@4.5 SMP8 using V6(Tracker) client, 560ti 2Gb. Win7 SP1, nVidia 301.42

OP is using a nVidia card: 560ti

He is pointing out that that the ppd (not the CPU usage) is dropping less on these WU's (p7149 & p7156) compared to p6940 when adding GPU P8009 (no change to smp8).

His comparison appears valid, therefore there may be something significantly different about these WUs that allows them to work better when part of a core is also running a GPU wu.

[gwildperson]

One thing that has annoyed me greatly in Folding is that adding a GPU client can
trash the ppd of the SMP client by 20-30%.

I guess the question is whether an NVidia GPU can trash the ppd of the SMP client by 20-30% or not. A lot depends on how many cores the CPU has and how much processing time a GPU uses. Certainly an ATI GPU can kill one CPU and with a Quad, trying to run 25% of a SMP project will slow up that WU by much more than 25%. If a NVidia GPU uses 3% of a CPU, it may slow down a SMP WU by much more than 3% but I doubt that's ever 25-30%.

Even though he has an NVidia GPU, he didn't specify if that was the system that was being referred to. I think we have to guess whether the fact that a GPU can trash SMP was something he was reporting about his system or something he heard reported about some other system. If it was his system, what projects were running and what percentages were being used by fahcore_xx and fahcore_yy when the system was being trashed?

[Simba123]

OP here,
I am referring to the system quoted - which you can see in detail herehttp://www.modsrigs.com/detail.aspx?BuildID=27676. The O/S is Win7 64bit SP1. With nVidia drivers 301.42
CPU is overclocked to 4.5
GPU is overclocked to 920/1840/2004

I have been folding for over a year now, and I can definitely confirm that adding a GPU task when simultaneously running an SMP task always reduces the ppd of the SMP task.

The GPU ppd remains constant - something to do with GPU scheduling having a higher priority than CPU/SMP - but the ppd of the SMP task is always reduced.

It appears that some SMP tasks are way more sensitive to the addition of a GPU task than others.
In general, I have found that adding in GPU consistently reduces the ppd of SMP by 20-30%, regardless of the GPU project type/number.

However, to confirm, that reduction when the SMP unit is a 7156/7149 series is only ~10%. Now WHY that is so I am not sure. I am assuming that something in the coding of the 7156/7149 series is different and it works better with GPU folding as well.

Given the efficiency gains, it is certainly something worth looking into. I know quite a few people who choose NOT to fold on their GPUs because the ppd drop on the SMP makes it hardly worth it. If the coders can work out what makes the 71xx different, and incorporate it into other projects, it may encourage more people to fold both SMP and GPU, which can only benefit the project overall.

Try for yourself if you doubt me.

I use the third party Tracker V2 program with untilizes the v6 client; I have also run the V7 client and observed the same results.

[Simba123]

Here is an example using my currently running projects.
Here is p6097 running for several hours without the GPU going. Tracker V2 uses the last 3 frames for
calculating the PPD, nothing else is running on the computer.


Here we can see what happens when the GPU is started. I let it run for more than 5 SMP frames so the
ppd calculation is based solely on when the GPU is running, not from when the SMP unit started.
As you can see, it dropped from 28105 down to 23757 a loss of 4348ppd or 18.3 %


Next time I get a p7149, if I'm around, I'll fire up the GPU and post what happens.

[Simba123]

@gwildperson, here is my resource usage at the time.

http://img854.imageshack.us/img854/1682 ... eusage.jpg

[Simba123]

Some more data for you.
start with P6984 SMP only


add GPU






Result



SMP only 28225
Add gpu, smp drops to 24135
Loss of 4090 ppd
or 14.5% (better than the last example)

[Simba123]

Here is our excellent P71xx series, in this case 7146

starting on it's own


Then add in GPU (same gpu PRCG as the last example)




In this case SMP only is 30694 ppd
add GPU and SMP drops to 28919 ppd
a loss of 1775 ppd
or 5.78%