i had forgotten that, although my web preferences for this host were set correctly, it is also using a cc_config.xml file with the line 1 in it. so i just added the 0 line to it, restarted BOINC, and the tasks that were running on the iGPU are now showing as "GPU missing." i no longer have to worry about the E@H server sending me intel iGPU tasks since the BOINC client no longer recognizes that GPU.
I lowered my iGPU clock from 1.35 to 1.30 GHz and could lower the voltage further (no power consumption measurements yet). This reduced the performance hit on my POEM WUs, surely because a lower clocked iGPU eats up less main memory bandwidth. Einstein performance also dropped a bit, of course, but those credits are easily reclaimed by POEM (much more than this, actually).
Edit: when I stop Einstein@iHPU now and no other tasks start, power draw drops by just 12 W. One could interpret this as "the iGPU only consumes 12 W", but sadly this would be wrong: when I stop Einstein, my nVidia GPU@ running POEM gains 5% GPU utilization and ~2% power consumption, which disturbs the measurement.
Something strange with this host. Can not download OpenCL BRP4G task while I opted out only of BRP5 tasks. And I'm unchecked "run CPU tasks for which GPU versions are available" long ago but just received BRP4 SSE tasks again. What it might be?
Pushing a bit further I have now collected this data for my HD4000:
1250 MHz @ -0.19 V
1300 MHz @ -0.16 V
1350 MHz @ -0.11 V
1400 MHz @ -0.00 V
With stock voltage presumably being ~1.10 V. Interesting to note here is that at 1.40 and 1.35 GHz I was getting invalid tasks when I lowered the voltage too far, with the rate of failures increasing if I pushed further. These are most probably caused by real errors in the ALUs, which otherwise don't affect the chips functionality in any visible way.
However, at 1.30 and 1.25 GHz I'm seeing almost no invalids any more. Instead, if I push too hard (or not hard enough, to be precise) I'm getting a Intel driver reset caused by some problem in rendering the Win 8 GUI. The GPU clock is reset to 1.15 GHz, one Einstein task hangs but recovers upon restarting it. This is far easier to diagnose and may allow to lower the voltage a bit futher if one day we can crunch on the iGPU without a display attached.
The voltage reduction has been pretty small going from 1.30 to 1.25 GHz so I'm getting close to diminishing returns here. Put POEM performance increased again ;)
Something strange with this host. Can not download OpenCL BRP4G task while I opted out only of BRP5 tasks. And I'm unchecked "run CPU tasks for which GPU versions are available" long ago but just received BRP4 SSE tasks again. What it might be?
perhaps you mistakenly changed your Einstein@Home primary (default) preferences, and your host has a set location of "home, work, or school?"...or perhaps the opposite happened (perhaps you mistakenly changed either your home, work, or school preferences, and your host is not set to one of those locations and just uses the primary (default) preferences)?
I do not understand, which type of current Einstein tasks may be processed on Intel GPU. Arecibo? Arecibo GPU?
Dear Bikeman!
Explain, please, a bit more, what do you mean under "OpenCL 1.2 capable driver, e.g 9.18.10.3165"?
Is "INTEL Intel(R) HD Graphics 4000 (1484MB) OpenCL: 1.01" (as BOINC pages show) compatible or not? And what is minimal driver version, if compatible?
I do not understand, which type of current Einstein tasks may be processed on Intel GPU. Arecibo? Arecibo GPU?
Dear Bikeman!
Explain, please, a bit more, what do you mean under "OpenCL 1.2 capable driver, e.g 9.18.10.3165"?
Is "INTEL Intel(R) HD Graphics 4000 (1484MB) OpenCL: 1.01" (as BOINC pages show) compatible or not? And what is minimal driver version, if compatible?
My HD4000 runs well. It processes Arecibo Tasks. Arecibo GPU tasks are not processed on Intel GPUs. I run two tasks at once and they take about 26 to 28 minutes to complete. 62.5 credits are awarded.
You need OpenCL 1.2! OpenCL 1.01 doesn't work. Download links can be found in this post.
my mistake...i got it all
)
my mistake...i got it all figured out.
i had forgotten that, although my web preferences for this host were set correctly, it is also using a cc_config.xml file with the line 1 in it. so i just added the 0 line to it, restarted BOINC, and the tasks that were running on the iGPU are now showing as "GPU missing." i no longer have to worry about the E@H server sending me intel iGPU tasks since the BOINC client no longer recognizes that GPU.
I lowered my iGPU clock from
)
I lowered my iGPU clock from 1.35 to 1.30 GHz and could lower the voltage further (no power consumption measurements yet). This reduced the performance hit on my POEM WUs, surely because a lower clocked iGPU eats up less main memory bandwidth. Einstein performance also dropped a bit, of course, but those credits are easily reclaimed by POEM (much more than this, actually).
Edit: when I stop Einstein@iHPU now and no other tasks start, power draw drops by just 12 W. One could interpret this as "the iGPU only consumes 12 W", but sadly this would be wrong: when I stop Einstein, my nVidia GPU@ running POEM gains 5% GPU utilization and ~2% power consumption, which disturbs the measurement.
MrS
Scanning for our furry friends since Jan 2002
Something strange with this
)
Something strange with this host. Can not download OpenCL BRP4G task while I opted out only of BRP5 tasks. And I'm unchecked "run CPU tasks for which GPU versions are available" long ago but just received BRP4 SSE tasks again. What it might be?
Pushing a bit further I have
)
Pushing a bit further I have now collected this data for my HD4000:
1250 MHz @ -0.19 V
1300 MHz @ -0.16 V
1350 MHz @ -0.11 V
1400 MHz @ -0.00 V
With stock voltage presumably being ~1.10 V. Interesting to note here is that at 1.40 and 1.35 GHz I was getting invalid tasks when I lowered the voltage too far, with the rate of failures increasing if I pushed further. These are most probably caused by real errors in the ALUs, which otherwise don't affect the chips functionality in any visible way.
However, at 1.30 and 1.25 GHz I'm seeing almost no invalids any more. Instead, if I push too hard (or not hard enough, to be precise) I'm getting a Intel driver reset caused by some problem in rendering the Win 8 GUI. The GPU clock is reset to 1.15 GHz, one Einstein task hangs but recovers upon restarting it. This is far easier to diagnose and may allow to lower the voltage a bit futher if one day we can crunch on the iGPU without a display attached.
The voltage reduction has been pretty small going from 1.30 to 1.25 GHz so I'm getting close to diminishing returns here. Put POEM performance increased again ;)
MrS
Scanning for our furry friends since Jan 2002
RE: Something strange with
)
perhaps you mistakenly changed your Einstein@Home primary (default) preferences, and your host has a set location of "home, work, or school?"...or perhaps the opposite happened (perhaps you mistakenly changed either your home, work, or school preferences, and your host is not set to one of those locations and just uses the primary (default) preferences)?
This is for overall settings,
)
This is for overall settings, not specially for Einstein. I pointed out to a log that shows something strange, that I can not understand.
RE: This is for overall
)
Ok. Waiting to drain down WU cache, then will try to reset the project just to see whether it helps.
I do not understand, which
)
I do not understand, which type of current Einstein tasks may be processed on Intel GPU. Arecibo? Arecibo GPU?
Dear Bikeman!
Explain, please, a bit more, what do you mean under "OpenCL 1.2 capable driver, e.g 9.18.10.3165"?
Is "INTEL Intel(R) HD Graphics 4000 (1484MB) OpenCL: 1.01" (as BOINC pages show) compatible or not? And what is minimal driver version, if compatible?
RE: I do not understand,
)
My HD4000 runs well. It processes Arecibo Tasks. Arecibo GPU tasks are not processed on Intel GPUs. I run two tasks at once and they take about 26 to 28 minutes to complete. 62.5 credits are awarded.
You need OpenCL 1.2! OpenCL 1.01 doesn't work. Download links can be found in this post.
So, factory drivers must be
)
So, factory drivers must be upgraded to version 1.2, yes?
Thank You, Maximilian!