CUDA error 702 failed to deallocate device memory

[mccsvt]

Hello,
I have seen that this topic has already been addressed, but the solution has never been fixed.
Beppe had advised to replace the drivers, but some did not fix it anyway.
Until yesterday I had Cinema 4d r14 with Octane 3.07 R2, driver 432.00 on GTX980 and I had the same error, so today I decided to buy the annual license of Cinema 4d r21 and install Octane 4.05 R7, installing new drivers 442.50 on the GTX980.
Why do I have the same error?
If it can help, the error is only on complex scenes where I use sky with hdri, instead if I use sky with sun everything works correctly.
I attach the logs.
Thanks
S

[bepeg4d]

Hi,
the c4doctanelog shows that the scene is too big for your machine, VRAM is finished, the scene is taking already ~2GB of Out-of-core, but maybe the assigned quantity is not enough.

It seems also that you are having an issue with timeout, are you using hig res textures for displacement?

What happens is that the Windows system instantly disable the GPU drivers, if it does not receive a signal after some seconds.
By default the TdrDelay is set to 2s, i.e. if one kernel call takes more than 2s, Windows kills the driver.
If you install the Standalone via installer, we change some Windows registry to better work with GPU Rendering.
To change the timeout, either run the Standalone installer or manually set the TdrDelay key to something like 10s or so. See here:

https://msdn.microsoft.com/en-us/librar ... s.85).aspx

In the Standalone installer we set these values:

WriteRegDWORD HKLM "SYSTEM\CurrentControlSet\Control\GraphicsDrivers" TdrLevel 3
WriteRegDWORD HKLM "SYSTEM\CurrentControlSet\Control\GraphicsDrivers" TdrDelay 10jh

Some Win 10 users have set the TdrDelay at 60, to solve their stability issues with CUDA.

ciao Beppe

[mccsvt]

Hi Beppe, happy to read you!
Thanks for the tips.
I had read about the TDR and therefore I had already made the change.
Now I have cleaned up the scene to save some VRAM which has dropped to about 1.6 Giga if I use sky + sun.
If I use Hdri, the VRAM increases to 2.5 Giga and I have about 700 Mb available ... so I will use Hdri only for final rendering.
But ... is it normal for HDRI (300 Mb) to increase VRAM to 1 Gb?
Thanks!
S

[bepeg4d]

Hi,
this is normal when the file is loaded, try to open the same HDR in Photoshop, and check the document size.
What makes the diffrerence is the resolution in pixel, and the pixel depth.
Try this:

• make a copy of the HDRI, and change the depth to 8bit, and save it in PNG or JPG, you should see a reduction of ~ a quarter in memory.
• make another copy of the HDRI, but this time reduce the resolution/size in pixel by 2x, again, you should see a reduction of ~ a quarter in memory.
• use the scaled HDRI in the Primary Environment node for lighting
• use the PNG/JPG version in the Visible Environment node for Backplate/Reflections/Refractions

ciao Beppe

[mccsvt]

Hi Beppe,
I did some tests.
I created two Sky: in the first I inserted the 8bit jpg image in the Visible Environment nod and in the second sky I inserted the hdr image scaled in the Primary Environment node. Then I did the test using the original 32-bit HDRI.
Comparing the values captured by the two different renderings, however, I know that the 2-sky method consumes more VRAM than the original HDR method.
Did I do something wrong?
Thanks

[bepeg4d]

Hi,
from your screenshots, there is 1GB of memory difference, ~200MB of VRAM, and ~850 of Out-of-core, but I cannot understand which one does have the visible environment, and which one does not

Anyway sorry, I was not accurate, you need to decrease the original HDRI resolution by 4x to see a difference.
Please, have a look at the following screenshots.
HDR 6k + white Visible Environment = 264MB:

HDR 1.5k + 6k 8bit Visible Environment = 178MB:

ciao Beppe