GPU contention

[IanKennedy]

I've been running the CPU version of Leela on my low spec laptop (hope to upgrade in the summer), running both training games and my own local testing matches, both with 1 thread each.

Just for laughs the other day I ran up the GPU version expecting it to not even find one but it reported an Intel Graphics HD 4400 which although no doubt lame by modern standards, still outperformed the CPU version.

I then merrily switched both my testing and training over to the GPU before noticing my fan had gone into overdrive and the nps had dropped back down although still outperforming the CPU version, and the total nps of the two instances still exceeded one running alone.

So now I've had to kill off GPU training and put that back on the CPU version but it does raise the question of how much of the GPU Leela consumes and how I can monitor it, other than the crude method of looking at nps?

By the way the initial GPU runs were unstable but after upgrading the driver from 10.18.10.3345 (2013) to 10.18.14.4414 (2016) the crashes went away and it sped up too.

[AdminX]

Try program like afterburner not to overclock, but just monitor. Plus it's free.

https://www.msi.com/page/afterburner

[IanKennedy]

Thanks, looks like I'm about 92% utilisation just running the one instance (either training or own local games) but 99% when I run both simultaneously.

[Milos]

Try program like afterburner not to overclock, but just monitor. Plus it's free.

https://www.msi.com/page/afterburner

Yea afterburner is a great tool for OC also, no matter which GPU brand you have. So even if you have ASUS, EVGA, Gigabyte or Gainward card (and not MSI) afterburner is probably better than anything native those manufacturers have.

[IanKennedy]

The problem with the fan noise is still apparent even with one instance now I've started paying attention to it and its starting to bug me. Its more obvious when LC-GPU moves and stops thinking as the drop off in noise is more abrupt. This is with 10' + 5" tc.