Very low credits for intel_gpu
2 Feb 2015 7:16:35 UTC
Topic 197956
(moderation:
BRP Arecibo on Nvidia GTS450 earns 1000 credits every hour.
Intel_gpu version on HD4600 earns 62 credits every 1,5 hour.
25 times worse!
However, according GPUs' performance the difference should be no more than 3-4 times (like I see in SETI).
Or BRP algorithm on Intels is so ineffective?
Very low credits for intel_gpu
)
It's actually a slightly different problem, with the run times.
The Intel GPU application is very sensitive to CPU usage: if your CPU is 100% utilised by 4 BOINC tasks, performance plummets. If you reduce the number of CPU tasks by one ("On multiprocessors, use at most 75% of the processors"), you will see much lower runtimes: my very similar host 8864187 turns out those 62-credit tasks every 11 minutes.
Be warned that there are currently some as-yet unexplained validation (accuracy) problems with HD 4600 GPUs and some driver versions. Keep an eye on the BRP4 Intel GPU app feedback thread at the top of the 'Problems and Bug Reports' area.
Thank you. It's really weird
)
Thank you. It's really weird app, doesn't eat cpu but requires one free core...
Yes, I recently started
)
Yes, I recently started crunching on i7-3770k/HD 4000 with CPU usage at 75%. I typically finish Einstein Intel GPU WUs in 12 - 12,5 minutes.
It's worth noting that the
)
It's worth noting that the credit granting is really "apples-to-apples" in all respects:
* the intel GPU OpenCL and ATI/AMD app versions are actually identical. The executable is just renamed. NVIDIA is different tho, of course (CUDA).
* the data that is crunched by the 'different' BRP4 apps is actually the same, just the "bundling" of the workunits is different: Intel GPU (and Android and ARM-Linux CPU ) tasks are "atomic" while the BRP4(G) workunits for NVIDIA and AMD GPUs consist of a "bundle" of several such tasks, and the factor in granted credit is this number of bundled tasks.
HB
RE: It's worth noting that
)
Thanks for the clarification.
I came to this conclusion a few days ago when I saw that my Nvidia card was receiving "bundles" of 16 files which had the same naming convention as those my iGPU is receiving. 1000 credits (what the Nvidia card is awarded) / 16 = 62.5, which is the credits earned by the iGPU per task.
Found the way to run IntelGPU
)
Found the way to run IntelGPU task at full speed without free cpu core. Just set priority to "Realtime" in Process Explorer, and Intel gpu usage became 96% instead of usual 15-20%. Run time decreased to 12 minutes.
RE: Found the way to run
)
Was the machine usable (normal desktop interaction, and typical daily tasks like browsing) with that setting?
Didn't mention any screen
)
Didn't mention any screen lags or so. Even on secondary display attached to intel video (primary on nvidia). That app actually doesn't use CPU (less than 1%), so everything ok.
I'll give it a try on my
)
I'll give it a try on my Haswell, which uses the iGPU to driver the monitor as well. But I'll need to wait for the uploads to clear, before I can download some more work.
RE: Found the way to run
)
Which task did you set to 'Realtime'? The einsteinbinary_BRP4.. ?
I tried it, but my win7 does not allow that. 'High' is the max I can set. And this increases the gpu usage just from ~50% to <60% average.
Could you please give more information?