Variable GPU load

[DiDu0815]

Do you see any chance to add separate options for variable load on GPU(s), similar to already existing CPU options (e.g. -l -p)?

[rschoene]

I assume that cublas<type>swap would be a suitable function? While this would still work on memory, it would not really use the available FPUs. Do you have a better suggestion (this will only be started after the AMD GPU integration)

PS: alternatively, we could not run any workload at all...
PPS: It could be hard cancelling the kernel and polling for the "low load" flag during synchronization. Any help and suggestions are welcome

[DiDu0815]

Unfortunately, I'm no way a good programmer so I have no idea about the best way to implement it. I am using FIRESTARTER for load tests to finetune cooling systems. I appreciated the functionality to vary the CPU load and was just curious whether that also would be possible for the GPUs as well.

[rschoene]

Thanks for your answer and sorry for my late replies. I have not much time to spare on this project.
Nevertheless, I keep that in mind. Here's a short explanation of the issue, so that you know what the issue is (and might be forgiving for me not having a straight-up answer).

While on the CPUs cancelling the high load is easy since the threads will monitor a certain variable every some tens to hundreds nanoseconds, it is not easy to cancel the workload that is running on the GPU. So the GPU threads could switch to a low-power routine after they finished one chunk of work. But this chunk of work takes quite some time, which means they might miss the deadline for switching to low (significantly) and sometimes even miss multiple successive low load phases.
Hence, I cannot just implement it that way, since a feature that does not exist is still better then a feature that does not work correctly.

I will have to rely on colleagues that are more knowledgeable when it comes to these things.