OcnaneBench dual GPUs ''out of memory' error

[v68]

Hello.
I've got two GPUs:
1. GTX 1060 6 Gb - plugged in PCIx16 3.0
2. GTX 750 Ti 2 Gb - connected via riser into PCIx1 2.0.


OctaneBench sees both cards but runs test only on single one.
When I select booth cards the app shows following:
GeForce GTX 750 Ti (2): and unspecified error occurred CUDA error 2 on device 1: out of memory -> failed to bind device to current thread.


If I restart the OctaneBench and choose test for 750ti - it runs but now 1060 has the same error "out of memory"

The Demo Octane for C4D also sees both cards but renders only on single one.
Whatever card I choose for render first - another one become unavailable for rendering - black viewport.
The log indicates following:

<<< Render failure detected!!! >>>
Please check render statistics to solve the problem.
MB:0/0 ST/MOV:0/0 Nodes:3 Tris:0 DispTris:0 Hairs:0 Meshes:0
Device:0 TotMem:6Gb rtData:0Kb film:0Kb geo:0Kb node:0Kb tex:0Kb unavailable:0Kb temperature:38
Device:1 TotMem:2Gb rtData:0Kb film:0Kb geo:0Kb node:0Kb tex:0Kb unavailable:121Mb temperature:41
VRAM used/free/max:0Kb/0Kb/0Kb
Free VRAM is too low!


Help me please to figure out how to bind both GPU for render.
Thank you.

Motherboard: z97m-ds3h (bios F7)
CPU: i7 4790k
RAM: 16 Gb

[bepeg4d]

Hi,
PCIx1 2.0 is too slow, please try to untick the Use for Tonemapping option in Preferences/Devices for the GTX 750TI.
Also note that 750ti has an average score of ~34, while we suggest a minimum of ~45 for a smooth experience:
https://render.otoy.com/octanebench/sum ... TX+750+Ti+
My suggestion is to try to use the 750ti for system/monitor/opengl only, and use the 1060 6GB for CUDA rendering only, in this way you can use the all 6GB for rendering.
If you enable also the 750ti, you cannot go over the 2GB limit, since the scene needs to be entirely loaded on each GPU.
ciao Beppe

[v68]

Hi,
PCIx1 2.0 is too slow

Not exactly.
I tested 750 ti on PCIx16 3.0 - the result is 32
On PCIx1 2.0 it is 31.03

try to untick the Use for Tonemapping option in Preferences/Devices for the GTX 750TI.

Done. Aint work.

Also note that 750ti has an average score of ~34, while we suggest a minimum of ~45 for a smooth experience:

I know that. And I got the 1060 to sum the processing power, which is 121 according to the Octabench
https://render.otoy.com/octanebench/sum ... GTX+750+Ti
This would work for me alright.

My suggestion is to try to use the 750ti for system/monitor/opengl only, and use the 1060 6GB for CUDA rendering only,
The point is to make them both to work on render.

Thanks anyway.

[bepeg4d]

Hi,
the PCI bus speed does not affect the rendering speed, that is roughly the same, as you have already seen, but the connection between GPU and CPU when the results have to be transferred, or when the scene is loaded from CPU to GPU.
Only 25% of more speed does not justify the lost of 4GB of VRAM, and the slow down in loading the scene, in my opinion.
The things could be completely different if you add another 1060 6GB, instead of GTX 750ti.
ciao Beppe

[v68]

The things could be completely different if you add another 1060 6GB, instead of GTX 750ti.

Yes, but if cannot yet bind both devices what the difference will be? Still will be working single GPU. As it is now.
First have to figure out what is happening.

[bepeg4d]

Hi,
with another 1060 6GB you can double the performance and have the complete VRAM.
Note that the quality of the riser is very important, and to correctly power up the bus.
ciao Beppe

[v68]

Hi,
Note that the quality of the riser is very important, and to correctly power up the bus.

Do you suppose that the riser could spoil the binding?

[bepeg4d]

Hi
I’m try to guessing, based on my forum experience.
ciao Beppe

[v68]

I’m try to guessing, based on my forum experience.

Did it happen before?

Alright, I'll try to connect another riser.

[v68]

Have tried another riser.
The same "Cuda error 2 out of memory"