Question Max allocation is 2.75 GB on a 11GB GPU

I have issues trying to use the GPU lightmapper in HDR,¨I have been trying for a few months now, with each new update but always fail, whatever I try to bake that is either a bit large or has more tha just a few meshes will automatically failback to CPU with the following message;

OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_MEM_OBJECT_ALLOCATION_FAILURE error executing CL_COMMAND_NDRANGE_KERNEL on GeForce RTX 2080 Ti (Device 0).

Or sometimes

OpenCL Error. Falling back to CPU lightmapper. Error callback from context: Max allocation size supported by this device is 2.75 GB. 3.00 GB requested.

Thing is, this is a 11GB GPU, how can it be restricted to so little VRAM ?

 

Hey @Weidz_,

The restriction is for a single allocation, not for the total memory used by the GPU lightmapper. Are you perhaps using a 4k atlas size?

Cheers!

 

I am trying to, indeed

 

seems like all lightmap sizes fall back to cpu.. (tested on 2018.4.x, and 2020.1.0a8, using gtx1080, rtx2080ti)

OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_MEM_OBJECT_ALLOCATION_FAILURE
OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_MEM_OBJECT_ALLOCATION_FAILURE error executing CL_COMMAND_NDRANGE_KERNEL on GeForce RTX 2080 Ti (Device 0).

 

still fails on 2020.1.0a9..
*although seems to bake if disable 9 out of the 10 light probe groups.. **nevermind, if failed halfway..


tested with this scene, adjusted all bake settings to quite minimum.
https://assetstore.unity.com/packages/3d/environments/urban/archvizpro-interior-vol-6-120489

2020.1.0a9 release notes mentioned these:

• GI: Improved memory allocation strategy for GPU lightmapper on Nvidia GPUs (Windows only).
• GI: Reduced GPU memory usage when baking lighting with the GPU lightmapper by using stackless BVH traversal.
• GI: Reduced memory usage for light probe occlusion baking when baking with the GPU lightmapper.
• GI: Compressed transparency textures for GPU lightmapper, 75% memory reduction for transparency textures by using rgba32 instead of floats.

some of the errors received with:

Code (CSharp):

1. OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_OUT_OF_RESOURCES error executing CL_COMMAND_READ_BUFFER on GeForce GTX 1080 (Device 0).
2. OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_OUT_OF_RESOURCES
3. OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_MEM_OBJECT_ALLOCATION_FAILURE error executing CL_COMMAND_NDRANGE_KERNEL on GeForce GTX 1080 (Device 0).
4. OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_MEM_OBJECT_ALLOCATION_FAILURE
5. OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_MEM_OBJECT_ALLOCATION_FAILURE error executing CL_COMMAND_COPY_IMAGE on GeForce GTX 1080 (Device 0).
6. OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_MEM_OBJECT_ALLOCATION_FAILURE error executing CL_COMMAND_READ_IMAGE on GeForce GTX 1080 (Device 0).

 

It seems to me a 4gb card cannot bake out a 4096 lightmap. Why not dice up the problem so it's tackled with multiple 512 lightmaps, then merge them after? Is this too difficult a job for Unity or is the answer simply to upgrade my 980 (which I can't afford to do) ?

 

ok, gpu bake works now (for archviz6 scene) using 2020.1.a15

@KEngelstoft what mainly affects to that baking performance mrays value?

(i mean if someone is getting 500mrays with 2070, seems that the scene complexity matters - but then again, one would expect to shoot "max amount" of rays from the device, regardless of what you are shooting at?)

 

@mgear I've upgraded to 2070S and the amount of rays will vary depending on options chosen.

 

tested with super simple scene and pretty minimal settings, still not going much over 220mrays.. i guess i'll test updating driver next.

Also baking still seems to fail, if cancel bake, then bake again, errors - need to restart unity to make it work.
(didnt happen with lower values though)

OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_MEM_OBJECT_ALLOCATION_FAILURE

OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_MEM_OBJECT_ALLOCATION_FAILURE error executing CL_COMMAND_NDRANGE_KERNEL on GeForce RTX 2080 Ti (Device 0).

 

Yep I used to restart if it fell back to CPU or it'd stay on CPU.

 

Nice to see this issue being solve. I am wondering if there is a May 2020 update for this as CPU versus my 1080Ti is really slowing down development?

 

So is it so hard for Unity to handle in 8 months or...? Same gpu mem problem in latest unity. Why is it that in 1 year Mr. F has singlehandedly made the Bakery light baking asset that just works and puts every single iteration of Unity's lighting to shame, and you still can't tile your renderer? I myself wrote 2 atlassers since then.

This is the kind of stuff this forum is riddled with that gives Unity such a bad name. 8 months come on, how many countries do you have offices in? How many people? And I can't even bake a quick demo scene on the gpu without an hour of reverse engineering unity and getting to a bad compromise.

"Progressive GPU" has been in "Preview" for how many YEARS now?

 

it has literally been 2+ years and this STILL is happening, what the hell is going on unity??
progressive has been in this unfinished state for almost half a decade already and it's still suffering from most nonsensical problems.

i'm over here trying just to bake a small scene and i'm having to deal with all of this stuff like searching up how to unlock open cl max memory JUST for a lightmap, it's insane.

 

it doesn't really change much as the lightmapper still uses only a quarter of your gpu memory, i have a 13 GB RTX 2060 GPU and it always says it runs out of memory despite not using even half of all the provided memory.

 

This is still an issue in 2023
I have a 3060ti that is only using a few hundred megabytes when I have a 8gb gpu

 

The issue is still there in 2024, with Unity 2022.3 using a rtx 2080ti. Seriously?