Message boards :
Number crunching :
GPU tasks only uses 50% load on GTX780TI
Message board moderation
Author | Message |
---|---|
Send message Joined: 28 Oct 10 Posts: 8 Credit: 80,345,093 RAC: 0 |
Can anyone explain why the GPU is not using 100% of available load. I have two GTX780TI installed and they only work at 50-55% load. Is there any settings to increase it? |
Send message Joined: 2 Dec 10 Posts: 12 Credit: 107,787,192 RAC: 0 |
Do you have a CPU core dedicated to each GPU as a supervisor? If your CPU is maxed out doing other work, then it will starve the GPU processing. As a quick test, you could temporarily suspend all CPU tasks and compare performance. Are the cards overheating and self-throttling perhaps? |
Send message Joined: 28 Oct 10 Posts: 8 Credit: 80,345,093 RAC: 0 |
It is no difference when I suspend all CPU tasks. The CPU load is about 2%. The cards are cold, about 50 Celsius and the fans are running at idle. |
Send message Joined: 8 May 09 Posts: 3322 Credit: 520,682,362 RAC: 33,626 |
It is no difference when I suspend all CPU tasks. The CPU load is about 2%. The cards are cold, about 50 Celsius and the fans are running at idle. First I think you need an "app_info.xml" file like this: <app_info> <coproc> <type>Nvidia</type> <count>0.5</count> </coproc> </app_info> I think that will tell your gpu to run two units at once, you WILL need to reserve a cpu core just for the gpu at that point or the units WILL run much slower. You can copy and paste the above lines into NOTEPAD in Windows, do NOT use a word processing program as it adds hidden stuff and the file won't run. Save the file as a 'txt type file' in the c:\program files\boinc\projects\milkyway folder as "app_info.xml" obviously leave off the quotes though. Stop and then restart Boinc and you should start running two tasks at once on each your gpus. Second are you actually using BOTH of your gpu's to crunch with? If not you may also need a "cc_config.xml" file to make that happen. The file would contain these lines: <cc_config> <options> <use_all_gpus>1</use_all_gpus> <skip_cpu_benchmarks>1</skip_cpu_benchmarks> </options> </cc_config> Use NOTEPAD and save it again as a 'txt type file' in the c:\program files\boinc folder as "cc_config.xml" again without the quotes. That way all projects you attach to will use both of your gpu's. You can either tell the Boinc Manager to read the config file or do a stop and restart of Boinc to make Boinc use the file. There are other config options such as: <exclude_gpu> <url>http://boinc.fzk.de/poem/</url> <device_num>1</device_num> [<type>NVIDIA</type>] </exclude_gpu> You can then put that inside the cc_config.xml file to use one gpu at one project and the other gpu at a different project. The above lines exclude the Nvidia gpu in my own system from the Poem project. The device number part is the key, each gpu will have a different device number, so changing it means you can tailor Boinc to work the way YOU want it to work. Here is a link to a website that gives most of the options available: http://boinc.berkeley.edu/wiki/Client_configuration#Application_configuration |
Send message Joined: 28 Oct 10 Posts: 8 Credit: 80,345,093 RAC: 0 |
Does anyone have an complete app_info file that works for all current applications that run under milkyway including cpu and GPU (Nvidia)? |
Send message Joined: 28 Oct 10 Posts: 8 Credit: 80,345,093 RAC: 0 |
Finally! After some searching on the forum I have finally high GPU load. I'm running 4 GPU units and 7 CPU units at the same time. The app_info.xml looks like this <app_info> <app> <name>milkyway_separation__modified_fit</name> </app> <file_info> <name>milkyway_separation__modified_fit_1.36_windows_x86_64__opencl_nvidia_101.exe</name> <executable/> </file_info> <app_version> <app_name>milkyway_separation__modified_fit</app_name> <version_num>136</version_num> <avg_ncpus>0.25000</avg_ncpus> <max_ncpus>0.510000</max_ncpus> <plan_class>opencl_nvidia_101</plan_class> <file_ref> <file_name>milkyway_separation__modified_fit_1.36_windows_x86_64__opencl_nvidia_101.exe</file_name> <main_program/> </file_ref> <coproc> <type>NVIDIA</type> <count>0.5000</count> </coproc> </app_version> <app> <name>milkyway_separation</name> </app> <file_info> <name>milkyway_separation_1.02_windows_x86_64__opencl_nvidia.exe</name> <executable/> </file_info> <app_version> <app_name>milkyway_separation</app_name> <version_num>102</version_num> <avg_ncpus>0.25000</avg_ncpus> <max_ncpus>0.510000</max_ncpus> <plan_class>opencl_nvidia</plan_class> <file_ref> <file_name>milkyway_separation_1.02_windows_x86_64__opencl_nvidia.exe</file_name> <main_program/> </file_ref> <coproc> <type>NVIDIA</type> <count>0.5000</count> </coproc> </app_version> <file_info> <name>milkyway_separation__modified_fit_1.36_windows_x86_64.exe</name> <executable/> </file_info> <app_version> <app_name>milkyway_separation__modified_fit</app_name> <version_num>136</version_num> <file_ref> <file_name>milkyway_separation__modified_fit_1.36_windows_x86_64.exe</file_name> <main_program/> </file_ref> </app_version> <file_info> <name>milkyway_separation_1.00_windows_x86_64.exe</name> <executable/> </file_info> <app_version> <app_name>milkyway_separation</app_name> <version_num>100</version_num> <file_ref> <file_name>milkyway_separation_1.00_windows_x86_64.exe</file_name> <main_program/> </file_ref> </app_version> <app> <name>milkyway_nbody</name> </app> <file_info> <name>milkyway_nbody_1.46_windows_x86_64__mt.exe</name> <executable/> </file_info> <app_version> <app_name>milkyway_nbody</app_name> <version_num>146</version_num> <max_ncpus>6.0000</max_ncpus> <plan_class>mt</plan_class> <file_ref> <file_name>milkyway_nbody_1.46_windows_x86_64__mt.exe</file_name> <main_program/> </file_ref> </app_version> </app_info> /Per |
Send message Joined: 8 May 09 Posts: 3322 Credit: 520,682,362 RAC: 33,626 |
Finally! After some searching on the forum I have finally high GPU load. I'm running 4 GPU units and 7 CPU units at the same time. You may want to experiment with leaving an extra cpu core free just so each gpu has it's own cpu core to use so it isn't waiting, obviously choose the faster way. ie only run 6 cpu units for a bit and check your rac and see if it goes up, if so then your gpu was waiting for the cpu to feed it. Since a gpu can do about 10 times more work than a cpu can, in the same amount of time, it makes sense to give the gpu the resources it needs. |
Send message Joined: 28 Oct 10 Posts: 8 Credit: 80,345,093 RAC: 0 |
How do I leave an extra CPU core free? Do I set <avg_ncpus> to a higher number e.g. 0.51 so I reserve 2 CPU for 4 GPU units? /Per |
Send message Joined: 18 Jul 09 Posts: 300 Credit: 303,562,776 RAC: 0 |
Boinc manager - advanced view - Tools tab - Computing Preferences - Processor usage - "On multicore systems use at most": put 90 in this box. |
©2024 Astroinformatics Group