we were just trying to run RealityCapture on a machine with 9 nVidia GTX 1080 GPUs.
Unfortunately, after 2 or 3 minutes into the process of creating the normal detail point cloud the machines switches off.
This is actually probably nothing related to RealityCapture only as we’re having this issue with other GPU-intense tasks as well (it’s probably the mainboard switching off due to some voltage issues).
BUT: what suprises us is the fact that nVidia Inspector is showing an increase of clock temperature and fan level for ALL of the 9 GPUs. But it should be only/max 3 that are in use, right?
I did a video with my phone (to be able to get the last frame before the reboot) of both nVidia Inspector and GPU-Z running while RC was processing the normal detail pointcloud. I attached screenshots of those. The GPU-Z instance at the top left is the one from the card that’s driving the monitors and which shouldn’t be used by any CUDA/OpenGL software.
Also, is there any way from stopping RC using GPU acceleration at all? Because right now we can’t use RCsince it won’t finish the processing and we’d have to wait to move it back to another machine. I couldn’t find anything regarding this in the settings.
Hi Holger Hummel
go to RECONSTRUCTION -> SETTINGS -> GPUs to use
and I highly recommend you to use just 1-2 GPUs and try to enable more of them after some testing…
Hi Vladlen
No, it is NOT because of this… For now these rules are still not used.
The 8 GPU setups are the troublemakers and it needs very careful tuning for stability…
“Each RC license is limited to 32 CPU cores and 3 GPU cards.”
May be this is a problem?
This is why i was writing in my original post that we’re surprised to see 9 GPUs being used.
RECONSTRUCTION -> SETTINGS -> GPUs to use and highly recommend to use just 1-2 GPU and try enabling more after good testing…
Thanks! This works. I was expecting those settings in the general RC preferences.
I’m running a test with 3 GPUs right now. Looking good so far
Will post an update soon.
No its NOT because of this… For now are this rules still not used.
Ah… so i guess this is during the current beta period?
The 8 GPU setups are trouble makers and it need very careful tuning fro stability…
This is something we figured ourselves already
With V-Ray we’re having stability issues from 9 GPUs onwards. Up to 8 works well so far.
This is why i was writing in my original post that we’re surprised to see 9 GPUs being used.
Usually this is only EULA, but possible add some limitation in code.
But if you buy 1 license for 3GPU but use it with 9, your licensed software not so far from pirated and can be punished.
Please do not think this way, as I have explained already few times, this rule is NOT being used like that, it is set because using more than 4 GPU is useless regarding speed-up of reconstruction. It is a “global” limitation, as photogrammetry uses quite heavy data transfers and PCI-E bus with its latency issues create problems with scaling - adding additional GPUs…
So for now there is NO limitation on GPU count per PC !
