If a computer has more than one GPU available to BOINC, all GPU tasks are assigned and executed on just one GPU.
I’ve noticed this on Windows, no idea about Linux.

Same here ...

Just noticed that the GPU which is selected has a "GPU Load" of 100%.
Even as it's only running ONE task.
An a "Memory Controller Load" of 100%.
The GPU temp is also very high.

Tried running on more than one GPU. No luck.
The tasks just pile up in one specific GPU.
But GPU-Z says that the tasks are distributed over ALL available GPUs.
I'm very confused over this behavior.

I'm new here, so I'm probably missing out on some important infos ?

Happy crunching !

[EDIT] As soon as the tasks are done, I will stop crunching and wait for instructions.

In the cc_config.xml file you must insert:

<use_all_gpus>0|1</use_all_gpus>

If 1, use all GPUs (otherwise only the most capable ones are used). Requires a client restart.

In reply to zioriga's message of 15 Mar 2026:

In the cc_config.xml file you must insert:

<use_all_gpus>0|1</use_all_gpus>
If 1, use all GPUs (otherwise only the most capable ones are used). Requires a client restart.

No, I don't think that’s the problem here. Both GPUs are available for BOINC. I don't have any problems with other projects. Even if that were true and BOINC were using only one GPU, it shouldn’t run two tasks on that single GPU.
I remember a similar issue occuring on Einstein and GPUGRID. It’s a bug/misconfiguration in the app’s code, not a client-side issue.

In reply to zioriga's message of 15 Mar 2026:

In the cc_config.xml file you must insert:

<use_all_gpus>0|1</use_all_gpus>

If 1, use all GPUs (otherwise only the most capable ones are used). Requires a client restart.

It doesn't work !

Thanks for the info, but of course I have the use_all_cpus set to 1.
The GPUs I have are all the same type.

I have been using these setups for years on, for example, EatH.

Thanks anyway !

Please, re-read my explanaition of the situation.

makracz is right — this was a bug in our app code. The application was hardcoded to use CUDA device 0 regardless of which GPU BOINC assigned to the task slot. So on multi-GPU machines, all tasks piled onto GPU 0 while the other GPU(s) sat idle.

This is fixed in v6.37 (deploying now). The app now reads the gpu_device_num from BOINC's init_data.xml and selects the correct GPU before doing any work. I rented a dual RTX 3090 instance to test this and confirmed each task runs on its assigned GPU.

To get the fix: Advanced > Update Project. Your tasks tab should show v6.37 for GPU tasks. If you still see an older version, try suspending and resuming the project.

Sorry for the trouble, and thanks for reporting it.

Indeed, it's working now.
I was wondering why?
Now I know.
Many thanks.

I'm wondering why you say it's fixed in 6.37 ?
In Applications it says 6.38 ?

So be it ...