OTOY Forums

Hi,

I was thinking about getting two RTX 4090s, but I just noticed by looking at their pictures that they don't have any NVLink connections. I then read that Nvidia dropped support for NVLink because of PCIe 5.0.
Theoretically, you could use the motherboard as the NVLink, but does it really work? Also, does this work for other 4000 series GPUs to double the VRAM?

Best regards,
Maxter

Hi, no double vRam here. Also for RTX 30xx series it works only for geometry.

They are working on streaming geometry and textures from an ssd (direct storage). I think that will be in the next big version of octane.

PolderAnimation wrote:They are working on streaming geometry and textures from an ssd (direct storage). I think that will be in the next big version of octane.

Are you talking about "Meshlets"? If so I'm really hyped for the next big update

SSmolak wrote:Hi, no double vRam here. Also for RTX 30xx series it works only for geometry.

Sadly So the NVLink is like the Out of Core function in Octane but much faster?

Maxter wrote:

SSmolak wrote:Hi, no double vRam here. Also for RTX 30xx series it works only for geometry.

Sadly So the NVLink is like the Out of Core function in Octane but much faster?

There is no NVLink at all for RTX 40xx series - even no such hardware slot for bridge. For RTX 30xx it works for geometry without or with very little speed impact.

PolderAnimation wrote:They are working on streaming geometry and textures from an ssd (direct storage). I think that will be in the next big version of octane.

Does this mean that we won't be limited by VRAM for our scenes?

PolderAnimation wrote:They are working on streaming geometry and textures from an ssd (direct storage). I think that will be in the next big version of octane.

What do you mean only on geometry. I was thinking of adding a 4090 but it seems having multiple 3xxx is better option ?

Creative_Sheep wrote:

PolderAnimation wrote:They are working on streaming geometry and textures from an ssd (direct storage). I think that will be in the next big version of octane.

What do you mean only on geometry. I was thinking of adding a 4090 but it seems having multiple 3xxx is better option ?

Memory pooling works for geometry only using RTX 30xx. So if most of you VRam is filled by textures you will not have benefit from NVlink. It will trtigger Out of Core instead.

How about delegating for rendering 3090 for Octane, while a 4090 for all other tasks; is this possible ?

What Nvidia has done by eliminated nvlink really breaks the desire to even own a 4xxx series card; unless you have multiple GPU.