Slow execution on multi-GPU machine
26 Sep 2013 20:41:38 UTC
Topic 197205
(moderation:
Hello,
yesterday i wanted to test running EAH on a machine i used for Bitcoin mining during the last year. However, i am seeing execution times of the Arecibo WUs of 8500 seconds compared to 2500 seconds seen on other machines. Albeit the GPU utilization being reported as 90%, the power consumption of the whole system only raises by 120 watts compared to idle - for a four GPU system! So i suspect that there is a problem in the configuration which causes this low performance.
This is the system:
AMD APU on standard motherboard
4x AMD HD7950 connected with 1x PCI-E extender cables and overclocked
Windows 7 32bit
BRP4G-opencl-ati version 1.39 gives me (as stated above) triple execution time. I already set the project to run 2 WUs per GPU, this increased the power consumption to 200 watts above idle. I will also try to run the GPUs with the stock clock rates tomorrow but i don't think this might solve it.
The only possible cause which i can't check is the PCI-E bandwidth requirement of the application. As written above, i use x1 extender cables which worked perfectly for mining Bitcoins because it needed no bandwidth. However, if EAH needs more than the supplied 250 MB/s on a x1 lane, this might lead to the performance degradation. Any ideas on this?
It's getting cold here with the system drawing 900 watts less power...
Regards,
Chris
Slow execution on multi-GPU machine
)
Einstein does require a lot of PCI-E bandwidth so limiting it to x1 would decrease performance, although I can't say by how much.
There has been some discussion of this in different threads in the "Cruncher's Corner" part of the forum.
There is also the "problem" with OpenCL apps requiring a full CPU core as support to get the best performance so if you are fully loading your CPU with other tasks try to free one or more cores as that should improve the performance of the OpenCL app.