use the <exclude_gpu> element

---

Your answer is posted at the project forum:

rebirther wrote:

https://srbase.my-firewall.org/sr5/forum_thread.php?id=6&postid=5765

The GPU apps can only run on device 0, it was not designed for multi-GPU.

That is incorrect, if I don't forbid any GPUs, It runs on both.

I'd double-check that theory against reality, if I were you.

BOINC will start two tasks, and give one the instructions to run on device 0, and the other instructions to run on device 1. BOINC Manager will show you that the instructions have been issued.

But what happens next is entirely dependent on the project's science application. If the application is not capable of reading and acting on the instructions, it'll do its own thing. I've seen that happen in the last few months, while a new application was being developed and tested at GPUGrid. They made a mistake in the code to implement device selection (or left it out entirely), and both tasks ran on device 0. We noticed, pointed it out, and they corrected it.

Part of the confusion arises from a change (many years ago) in the way BOINC passes the device instruction. When GPUs were first introduced, it was a very simple parameter on the command line: as GPUs got commoner and more diverse, that was changed to a more complex structure in a file called init_data.xml. Not every project got the memo.