single 8970M now appears and runs (a bit faster) as GPU0 and GPU1
12 Mar 2019 2:05:36 UTC
Topic 218374
(moderation:
It's not a bug but I'd like to know what has happened. No rush.
I have not used my laptop for E@H for over a year. I Installed (from scratch) the latest client and was a bit surprised.
Laptop has an integrated intel GPU and a discrete AMD GPU. For some strange reason the AMD is now discovered as two separate GPUs and there are two interdependent WU running on each one. Intel has its own WU as well.
For a second, I thought that E@H client somehow can use it as two separate units but it effectively shares the same one. When running in parallel each copy finished one WU in 34 mins. When only one is allowed to run, it does it in 18+ min. So it showed even a slight speed up when running on 'separate' GPUs. WUs were correctly validated.
I understand that I can turn off the "second" ATI GPU in cc_config.xml but do not have a reason to do it - its a bit faster now.
Is it by design or a funny problem with OpenCL device discovery in "7.14.2 for windows_x86_64" which somehow ended as a benefit?
Also GPU RAM is reported incorrectly. The AMD has 4GB.
Here's the OpenCL output from the log file:
OpenCL: AMD/ATI GPU 0: AMD Radeon(TM) HD8970M (driver version 2766.5, device version OpenCL 1.2 AMD-APP (2766.5), 3072MB, 3072MB available, 2304 GFLOPS peak)
OpenCL: AMD/ATI GPU 1: AMD Radeon(TM) HD8970M (driver version 2766.5, device version OpenCL 1.2 AMD-APP (2766.5), 3072MB, 3072MB available, 2304 GFLOPS peak)
OpenCL: Intel GPU 0: Intel(R) HD Graphics 4600 (driver version 20.19.15.5058, device version OpenCL 1.2, 1630MB, 1630MB available, 208 GFLOPS peak)
OpenCL CPU: Intel(R) Core(TM) i7-4800MQ CPU @ 2.70GHz (OpenCL driver vendor: Intel(R) Corporation, driver version 5.2.0.10094, device version OpenCL 1.2 (Build 10094))
Best
Pawel
BeemerBiker had the same
)
BeemerBiker had the same issue with 2 cards, BOINC saw 4 cards. See this thread on the BOINC forums
https://boinc.berkeley.edu/forum_thread.php?id=12807&postid=89777#89777
Thanks MMONNIN! Will attach
)
Thanks MMONNIN!
Will attach to the other thread.
The linked BOINC thread is
)
The linked BOINC thread is primarily concerned with removing the phantom GPU occurrence (believed to be caused by multiple driver instances).
If the enquirer finds that parallel processing using the duplicate driver increases efficiency, it would be better to refer to the project's officially-supported parallel processing route via https://einsteinathome.org/account/prefs/project.
Richard Haselgrove wrote:The
)
Originally I wanted only to remove the duplicate but the efficiency increase was a bit strange. Anyway, it might be solely connected to E@H client but please re-post this question anywhere if it sounds interesting.
BTW The link in the quote leads to my project preferences. It is the correct one to the parallel processing route?
pzajdel wrote:BTW The link in
)
Yes, that was quite deliberate - that is where you can control the behaviour of your own devices.
Scroll down to the bottom of the page, and see the separate sections starting "GPU utilization factor of ..."
If you set a task-type to use 0.5 (half) of a GPU, two tasks can run at once.
OK So if I get it right even
)
OK So if I get it right even with one GPU detected I could have mimicked my current "ghost" setup with 0.5.
I'll try to play with it. Thanks for the tip.
pzajdel wrote:OK So if I get
)
You can also configure any of the projects with a app_config.xml file in the project directory. The Einstein one below does the same thing and allocates 1 cpu resource for each of the WU. If you wanted to run 3, change the 0.5 to 0.33. I take the lazy way of finding out the names of the applications in a project by just inserting any bogus name. BOINC will print an error message in the EVENT LOG that tells what the acceptable names are.
<app_config>
<!--
comments
-->
<app> <name>hsgamma_FGRPB1G</name> <gpu_versions> <cpu_usage> 1.0 </cpu_usage> <gpu_usage> 0.5 </gpu_usage> </gpu_versions> </app>
</app_config>
OK I am still at the two GPUs
)
OK I am still at the two GPUs config.
I changed FGRP (in fact all projects) utilization to 0.5 through the webpage but it still runs 1 WU per "core". I expected it to run 2x2.
I saved and updated the project to run from the web settings. Other settings changes as expected but changing 1.0 to 0.5 did not cause BOINC to double the load. I do not have much time to work on it but will have a look.
Thanks for all tips.
pzajdel wrote:... did not
)
As Richard mentioned, the setting doesn't apply until new work is fetched. Clicking 'update' without triggering a work fetch isn't enough. To have an immediate effect, just increase your work cache locally through BOINC Manager so that your client asks for new work. Then restore the setting if you wish.
EDIT: Just be aware that running four concurrent tasks on your particular GPU (which is really what you are trying to do apparently, if BOINC still sees two GPUs) may make things far worse or even give rise to compute errors or system crashes. I'm pretty sure your 'best' result will come from just two concurrent tasks, however you achieve that.
Cheers,
Gary.
Gary Roberts wrote:pzajdel
)
Good point. You are much better at your replies than I am.
Reading the config files will change the behavior, but all the WU already downloaded will not show the new CPU/GPU percentage. The newly fetched WU will show the new fractional from the app_config.xml file.