1 card good, 2 cards bad? Tweet

[Rik]

Hi Paride

Finally found some time to investigate...

I'd been using a pair of 2080ti's without a hitch for years. Maximum parallel samples and never ran out of Vram.
Upgraded to latest version of Octane and installed latest Nvidia Studio driver.

Since then, very unstable and flakey regarding Vram. Same old models/textures/ resolutions etc so maybe it's the version of Octane or the new driver?

Tried just one 2080ti in the rig, rather than 2 and it worked perfectly in various PCI slots and with different connectors from the PSU.
Replaced with the other card and it also worked perfectly so pretty sure the cards aren't the problem.

So they work fine on their own but something goes tits up when try and use both.

They're 11GB cards and GPUZ shows max use of only 5GB so seems like it's not actually a Vram issue.
Power draw on the cards doesn't get above 76W so it's crashing before they even start working.

PSU is 1200W Gold

Any ideas why it now wont work with 2 cards but is fine with 1?


If I can work out how to roll back the Nvidia driver I'll give that a go.

Edit: Updated to the very lates Nvidia Studio driver. Still no use...

CUDA error (2) on device 1: out of memory
-> failed to allocate device memory
device 1: failed to allocate ai light updates buffer (3)
OptiX error for device 0: [ACCEL] Error launching command list (CUDA error string: out of memory, CUDA error code: 2)
Failed to build acceleration structure
OptiX API error for device 0: Device out of memory
OptiX error for device 0: [ERROR] "sbt->raygenRecord" is null
OptiX API error for device 0: Invalid value
device 0: RTX picking failed

[paride4331]

Hi Rik,
Last time, I believe your issue was due to an incorrect GPU Headroom setting.
Please try setting the GPU headroom to 512 MB instead of 100
Regards
Paride

[Rik]

2 cards 8 parallel samples, fail

1 card 32 parallel samples, all good

Hi Paride

I did try that but it still crashes.

Don't understand why it will now only work properly with only 1 graphics card installed?
I can run a high res render, maxed out parallel samples and minimum GPU headroom no problem at all with 1 card.
When I put the other card back in it fails.

[paride4331]

Hi Rik,
This could make sense if you disable the GPU (in the Octane device panel) that is connected to the monitor, as it uses VRAM for Windows, 3ds Max, Chrome, Photoshop, and so on.
Thus, it has less available VRAM, and when you activate it for GPU rendering, it also cuts into the VRAM of the other GPU, causing Octane to go out-of-core.
Regards
Paride

[Rik]

Maybe, if it was crashing with only 1 card. But it works fine with 1 card, which is connected to the monitors.

How can it work with 1card but not 2?

[paride4331]

Maybe, if it was crashing with only 1 card. But it works fine with 1 card, which is connected to the monitors.

How can it work with 1card but not 2?

Maybe, if it was crashing with only 1 card. But it works fine with 1 card, which is connected to the monitors.

Are you 100% sure?

Anyway, to verify this You could perform a quick test using both GPUs, setting 'parallel sample' to 1 and 'Max tile sample' to 1. Also, please set the headroom to 512 MB.

[coilbook]

Maybe, if it was crashing with only 1 card. But it works fine with 1 card, which is connected to the monitors.

How can it work with 1card but not 2?

Maybe, if it was crashing with only 1 card. But it works fine with 1 card, which is connected to the monitors.

Are you 100% sure?

Anyway, to verify this You could perform a quick test using both GPUs, setting 'parallel sample' to 1 and 'Max tile sample' to 1. Also, please set the headroom to 512 MB.

Our slaves always get these errors while cards on the master are fine.

[Rik]

100% sure Paride.

Here's the results from latest test as requested.

Please remember, using either card by itself works perfectly with settings maxed out - 32PS, 64 MTS, 100mb GPU headroom.

Previously, this was also the case when using both cards.

Only things to have changed are the version of Octane and the Nvidia Driver.

2 cards 1 parallel sample, 1 Max tile samples, GPU headroom 512mb, fail

[paride4331]

Hi Rik,
Could you try disabling RTX and denoiser, and test with or without out-of-core?
Also, would you share the device panel with the following conditions: only the first GPU, only the second GPU, both GPUs (first GPU and second GPU)?
Regards
Paride

[Rik]

2 cards 1 parallel sample, 1 Max tile samples, GPU headroom 512mb, Out of core enabled, fail

Hi Paride, thanks for looking into this. Good to not feel alone!
RTX and denoiser were disabled for the previous failed render and this one.

This one also now has out of core enable, 12GB

Seems like the GPUs aren't getting the model fully loaded as when I use a single card it's shows about 8.5Gb whereas with 2 cards they don't load more than 4.13Gb each. That's about 1/2, almost as though Octane is sharing the model between the 2 cards.
Also, they don't get around to drawing any power. This can also be seen on the GPUZ panels.

Not sure how to show both GPUs at once in Task Manager? Though there's a load of info for them in the GPUZ panels.