GPU PPD drops when adding more

[xl3b4n0nx]

I have a GTX 1080 in my machine that normally gets 700K ~ 800K on its own. I recently added 2 Quadro 2000s and they can do around 100K a piece . The problem is that when I run all 3 the GTX 1080 drops to around 100K PPD. I am running a Ryzen 2600X so I am not short on CPU cores and each GPU gets one. When I tried to setup all of the GPU slots it complained that I needed to manually set the OpenCL index. I have them set at 0, 1 and 2. If I remove the two Quadro 2000 slots the 1080 goes back to its normal performance. Does anyone know why this happens? Am I configuring FAH Client wrong? I really want to have all 3 running full tilt.

[Gorgon]

Could it be an issue with the number of PCIe lanes allocated to the slots?

 

Under Windows a 1080 would need between 4-8 lanes while the Quadros could likely get by with just 4 lanes. Rearranging the cards might help.

 

Another issue is that each NVidia slot should also have one core/thread reserved for each card by default. I believe Windows should do this for you but under Linux it is possible if you are using the additional CPU cores/threads for folding or BOINC there might not be enough cores/threads to keep the GPUs fed.

 

The FAHClient is generally pretty dumb and relies on the underlying OS for resource allocation.

[xl3b4n0nx]

1 hour ago, Gorgon said:

Could it be an issue with the number of PCIe lanes allocated to the slots?

 

Under Windows a 1080 would need between 4-8 lanes while the Quadros could likely get by with just 4 lanes. Rearranging the cards might help.

 

Another issue is that each NVidia slot should also have one core/thread reserved for each card by default. I believe Windows should do this for you but under Linux it is possible if you are using the additional CPU cores/threads for folding or BOINC there might not be enough cores/threads to keep the GPUs fed.

 

The FAHClient is generally pretty dumb and relies on the underlying OS for resource allocation.

I checked in GPUz and the 1080 is on a x16 link and the Quadros each have a x4 link. The link arrangement happens at boot and if I remove the slots for the Quadros, the 1080 goes back to full performance. I don't think PCIe allocation is the issue here. I am running the latest Nvidia driver which is too new for the Quadro 2000s. Also, they don't support openCL. That may be the issue.

 

My CPU folding dropped a thread for each GPU added so there is no shortage in CPU power for them .

[Metallus97]

Try to disable COU folding for a day or so and see if that helps. 

not so drastic way of doing things: go to advanced settings and manually put in your CPU slots and do yourCores-2xgupInstaces. So then there are 2 cores for each CPU. I know the GPU core21 is single Thread, but maybe this reduces the overhead. Possibly due to Hyperthreading? (Dunno how AMD calls it LOL)

[xl3b4n0nx]

I have changed out the GPUs. I am now running a GTX 1080 and a Quadro K4000. I have disabled CPU folding and I am having the same issue. The GTX 1080 is GPU 1 according to task manager. The Quadro is running at full speed (~700K-800K PPD) and the 1080 is running at ~35K PPD. I have no idea why....... 

[Ithanul]

Hmmm, I say pause one of the cards and look at which card is getting a load through either Windows Manager or a 3rd party application and see if that matches up to the one shown in F@H Control.  Sometimes F@H Control can be weird and show the incorrect label for the GPU that is folding.

 

Had that problem back when I had my 960 with a 980.  The 960 would show up as the 980 and the 980 would show up as the 960 in F@H Control.

[xl3b4n0nx]

I've had that happen too. At the moment they are labeled correctly. I had them configured with the cuda and openCL indexes set manually (Quadro 2000 and 1080 setup needed that). I have changed it to -1 for those on both GPU slots and it seems to be behaving. It usually takes about 10 minutes for the PPD to drop so we will see what happens.

[xl3b4n0nx]

Now the quadro took a nose dive on points. This is really frustrating.