High CPU usage with Sapphire R9 270X GPUs

[teodor_dj]

Hi,
my config is as follows:
CPU: Intel Pentium G3220 2cores 3.0GHz
MB: Asrock H81 Pro BTC
RAM: Adata 4Gb DDR3 1600MHz
HDD: WD Blue 320GB
GPUs: Sapphire R9 270x - 6 pcs.
PSU: EVGA SuperNOVA NEX1500 Classified
OS: Windows 8.1 x64
AMD Drivers 14.4
FAHClient v7.4.4

I've been using this config for mining without any problems. Recently started folding and I'm having a lot of problems. The main is that CPU is at 100% load constantly and GPUs performance suffers from this. I get 45000-65000PPD per card and the system is unstable and freezes sometimes. For the experiment, I changed the CPU with Intel i7 4770 (4 cores, 8 threads) and guess... no change at all. CPU load 100%, same problems. Then tried to lower the count of GPUs and at 4 pieces everything is perfect - CPU load 15-20%, and 75000-85000PPD per card.
So, anyone have same problems? Is this a FAHClient bug, AMD drivers bug or anything else?
Thanks!

[bruce]

Welcome to foldingforum.org, teodor_dj.

I'm going to assume that the task that's keeping the CPU busy is the FahCore* rather than FAHClient or FAHControl. If that's not accurate, let me know.

The developers of OpenCL drivers seem to like to write them to keep one CPU busy per GPU. Personally, I think it's a bad design choice, but I don't know enough about developing drivers to know the trade-offs. Folding@home is at the mercy of AMD (and NVidia) to develop the best drivers that they can.

[teodor_dj]

I'm going to assume that the task that's keeping the CPU busy is the FahCore

- that's right, when all six GPUs are running simultaneously, each FahCore task takes about 15-25% CPU processing power and thus CPU load hits 100%, obvously all six FahCores combined need more processing power than CPU can provide. That's OK, but when only four GPUs are running, each FahCore task consumes only 1-3% CPU power, not even speaking of taking a whole core by one GPU! I've tested this also on another PC, and results are the same - 1-2 % CPU usage per GPU(FahCore). That's weird enough.

Now I'm running four GPUs on the aforementioned rig, and the other two - on my PC. That is my solution for now. Any ideas are welcome! Thanks!

[P5-133XL]

For Core_17 WU's, when the client first starts, the CPU load on Core_17 for each AMD GPU will be very high but this should last for only a few minutes. With 6 cards, it would not surprise me that it gets to 100%. This is normal, for Core_17 sets up the WU using the CPU before it starts processing it inside the GPU. Once the %GPU usage goes high (95%-100%) The CPU usage should drop to 1-3% except for a periodic spike that should last only a few seconds at a time.

What I'm speculating is that with the 6-GPU running you measured the %CPU when they first started and with the 4-GPU setup the measurement was slightly later after the initial startup was finished.

[teodor_dj]

What I'm speculating is that with the 6-GPU running you measured the %CPU when they first started and with the 4-GPU setup the measurement was slightly later after the initial startup was finished.

- unfortunately, this is not the case.
My findings refer after the initial startup process. Looking retrospective, I would add that in 6-GPU configuration, all GPUs paused - CPU usage is a couple %. Then start GPU Number_1 and wait.. Initially CPU usage goes high(as you describe), after a couple of minutes drops to 15-25%. This is too high. With 4-GPU configuration, after the initial startup process, CPU usage drops to 1-2-3% . Strange situation. It appears, that even not folding, the fact, that 6 GPUs are connected to the motherboard, aggravates CPU usage for just one GPU folding. Continuing to start next GPUs, at GPU Number 4, CPU usage already hits 100%, measured after the initial startup process.
To summarize, for 4 GPUs running:

4-GPU Configuration - CPU usage 12-20%
6-GPU Configuration - CPU usage 100%

[7im]

With only one x16 lane, and five x1 lanes, the GPUs are probably fighting for bandwidth/data. CPU feeds the data. That's why we recommend 1 CPU core per GPU, not 3 GPUs per 1 CPU.

x1 lanes are known to bottleneck GPUs while folding by at least 10%. The faster the GPU, the worse it gets.

[bruce]

You've done some research to uncover how the %CPU varies. How about comparing the CPU used with bus speed.

For example, do the 16x lanes settle down to a smaller percentage than the 1x lane?

If you combine 2 16x lanes, does it use the same as, say, a 1x and a 2x?

When using GPUs on limited bandwidth slot, Very little is documented.

[teodor_dj]

Due to using extenders, all GPUs are at x1 PCIe speed. So they have equal bandwidth. Two GPUs running on my PC are on x16 lanes, no difference at all - 1-2% CPU usage per GPU, but slightly higher PPD.
Just to remind, when only four GPUs are installed, all of them on x1 lanes and ratio is 2GPUs\1CPU core (Pentium G3220), there isn't any problem - CPU usage is 12-20% for all tasks. When all six GPUs are installed, but only four of them are running, CPU is at 100%. Swapping CPU with Core i7 4770( 4 cores with hyper-threading, 3.4-3.9GHz) doesn't help at all, CPU is at 100% mark again. i7 4770 is a rather powerful CPU, it's difficult to accept that it cannot manage four running GPUs(as recommended 1 CPU core per GPU).
Finally, after a considerable amount of time spent, I gave up trying to dig any further. Now running 4 GPUs on the rig and 2 GPUs on my PC.

[PantherX]

May I suggest that you post over at the official AMD Forum (http://forums.amd.com/) to see what their response is to 6 GPUs in a single system (do note that F@H doesn't use CFX for folding)? While 4 GPUs seems to be supported in some benchmarks and games, having 6 GPUs might be an untested and/or unsupported configuration by AMD so may contain bugs. If you have indeed discovered a bug and AMD is willing to fix it, it may be advantageous to other multi-GPU folders in the future.

[teodor_dj]

Thanks, will do!