Cycles - Out of Memory Bug #45629

New Issue

Vlad Mafteiu Scai · 2015-07-31T15:39:01+02:00

Vlad Mafteiu Scai commented

2015-07-31 15:39:01 +02:00

System Information
Windows 10, 2x 780 GTX

Blender Version
Broken: 2.75a Official
Worked: 2.74 Official

I get this error when attempting to render
CUDA error: Out of memory in cuLaunchKernel(cuPathTrace, xblocks , yblocks, 1, xthreads, ythreads, 1, 0, 0, args, 0)"

I have 3 GB of VRAM available and they are barely filled. What's going on?

Exact steps for others to reproduce the error

One way to get the error is to viewport render then Ctrl+Z to disable it, then press F12 to tile render & after that progressive render again. When attempting to render in the viewport after this error appears. However after I imported some mesh in the scene, error appears each time and now I can't render anything.

Check attached file.

Room 2 .blend

**System Information** Windows 10, 2x 780 GTX **Blender Version** Broken: 2.75a Official Worked: 2.74 Official I get this error when attempting to render CUDA error: Out of memory in cuLaunchKernel(cuPathTrace, xblocks , yblocks, 1, xthreads, ythreads, 1, 0, 0, args, 0)" I have 3 GB of VRAM available and they are barely filled. What's going on? **Exact steps for others to reproduce the error** One way to get the error is to viewport render then Ctrl+Z to disable it, then press F12 to tile render & after that progressive render again. When attempting to render in the viewport after this error appears. However after I imported some mesh in the scene, error appears each time and now I can't render anything. Check attached file. [Room 2 .blend](https://archive.blender.org/developer/F216362/Room_2_.blend)

Vlad Mafteiu Scai commented

2015-07-31 15:39:01 +02:00

Changed status to: 'Open'

Sergey Sharybin was assigned by Vlad Mafteiu Scai

2015-07-31 15:39:01 +02:00

Vlad Mafteiu Scai commented

2015-07-31 15:39:01 +02:00

Added subscriber: @00Ghz

Vlad Mafteiu Scai commented

2015-07-31 15:53:07 +02:00

After some experiments I discovered that it has something to do with adding an Environmental texture to Background.

Simple test scene was a Suzanne with Subdiv Level 2, a plane and a 8k jpeg image for the Env Map. Also a MIS with 256 resolution.

There is no way that little data can fill over 3GB of VRAM.

Not running any other GPU intensive apps either.

After some experiments I discovered that it has something to do with adding an Environmental texture to Background. Simple test scene was a Suzanne with Subdiv Level 2, a plane and a 8k jpeg image for the Env Map. Also a MIS with 256 resolution. There is no way that little data can fill over 3GB of VRAM. Not running any other GPU intensive apps either.

Thomas Wilmot commented

2015-08-01 04:40:46 +02:00

Added subscriber: @ThomasWilmot

Thomas Wilmot commented

2015-08-01 04:40:46 +02:00

I think it has something to do with direct x 12 Unified Memory (Stress: I think)

I have a 4k monitor with 2 2gb 680's, and due to my cards struggling to run the desktop environment, rendering with the gpu on even the simplest scenes wasn't viable on 8.1. However, now that I've moved to Windows 10, I can now once again render on my GPU, however using anything over about 800mb of RAM causes this issue.

Also for me, if you render in the render window, only about a quarter of the tiles render.

Hope this helps :)

I think it has something to do with direct x 12 Unified Memory (Stress: I think) I have a 4k monitor with 2 2gb 680's, and due to my cards struggling to run the desktop environment, rendering with the gpu on even the simplest scenes wasn't viable on 8.1. However, now that I've moved to Windows 10, I can now once again render on my GPU, however using anything over about 800mb of RAM causes this issue. Also for me, if you render in the render window, only about a quarter of the tiles render. Hope this helps :)

perfection cat commented

2015-08-01 04:58:28 +02:00

Added subscriber: @sindra1961

perfection cat commented

2015-08-01 04:58:28 +02:00

Blender does not support windows10 officially.
I think that you should report it if the problem seems to still occur when it came to be supported.

Blender does not support windows10 officially. I think that you should report it if the problem seems to still occur when it came to be supported.

Thomas Wilmot commented

2015-08-01 05:25:39 +02:00

Regardless of whether it is or isn't officially supported, I believe that it's still a good idea to report bugs that get encountered, that way the developers know that the bug exists.

perfection cat commented

2015-08-01 06:21:12 +02:00

@ThomasWilmot:
Do you know this?
NVIDIA's Latest Windows 10 Drivers: Still Too Unstable For Primetime

@ThomasWilmot: Do you know this? [NVIDIA's Latest Windows 10 Drivers: Still Too Unstable For Primetime ](http://www.dailytech.com/NVIDIAs+Latest+Windows+10+Drivers+Still+Too+Unstable+For+Primetime/article37267.htm)

perfection cat commented

2015-08-01 06:31:34 +02:00

Blender uses Cuda 6.5, but it does not support windows10.

Vlad Mafteiu Scai commented

2015-08-01 07:05:42 +02:00

Which has been launched btw. Maybe it’s time to support it? :)

Will check if older drivers fix this issue.

And I don’t have mobile GPU’s either.

Which has been launched btw. Maybe it’s time to support it? :) Will check if older drivers fix this issue. And I don’t have mobile GPU’s either.

Thomas Dinges commented

2015-08-01 10:50:57 +02:00

Added subscriber: @ThomasDinges

Thomas Dinges commented

2015-08-01 10:50:57 +02:00

To clarify, does it work with Blender 2.74 on Windows 10? Or was that tested on a older Windows OS?

Setting to Low for now, likely a driver problem and nothing we can fix right now.

To clarify, does it work with Blender 2.74 *on* Windows 10? Or was that tested on a older Windows OS? Setting to Low for now, likely a driver problem and nothing we can fix right now.

Martijn Berger commented

2015-08-02 16:35:20 +02:00

Added subscriber: @MartijnBerger

Martijn Berger commented

2015-08-02 16:35:20 +02:00

Are you using an SLI bridge between the 2 cards ?

How much vram is in use before you open up blender ? ( use task manager thingy or GPUZ )

Are you using an SLI bridge between the 2 cards ? How much vram is in use before you open up blender ? ( use task manager thingy or GPUZ )

Julian Eisel commented

2015-08-02 16:53:16 +02:00

Added subscriber: @JulianEisel

Julian Eisel commented

2015-08-02 16:53:16 +02:00

Seems like we are awaiting responses from author - setting to incomplete prio.

Vlad Mafteiu Scai commented

2015-08-02 17:17:21 +02:00

Was tested on Win 8.1 for 2.74

Not using SLI bridge.

Around 300-400MB VRAM used if blender is closed.

I installed a GPU driver from March and the error seems to be less frequent now but not gone.

Was tested on Win 8.1 for 2.74 Not using SLI bridge. Around 300-400MB VRAM used if blender is closed. I installed a GPU driver from March and the error seems to be less frequent now but not gone.

Julian Eisel commented

2015-08-02 17:20:22 +02:00

Aaand back to low.

Martijn Berger commented

2015-08-02 20:35:50 +02:00

can you see how much ram rendering the default cube is using ?

measure when blender is open but not rendering.. then when it is rendering a 10000 sample default cube.

I think there might be some uvm crap doubling the (hidden) initial memory use for starting a kernel.

can you see how much ram rendering the default cube is using ? measure when blender is open but not rendering.. then when it is rendering a 10000 sample default cube. I think there might be some uvm crap doubling the (hidden) initial memory use for starting a kernel.

Vlad Mafteiu Scai commented

2015-08-02 21:17:46 +02:00

Sure.

What kind of render? Viewport or tiled or doesn’t matter?

Also the memory report in the info bar is bugged. It doesn’t seem to want to reset. It keeps reporting more and more memory being used for each repeated action. Even selecting/deselecting or repeated renders seem to get the reported memory to over 6GB....

Any idea how to get an accurate report?

Sure. What kind of render? Viewport or tiled or doesn’t matter? Also the memory report in the info bar is bugged. It doesn’t seem to want to reset. It keeps reporting more and more memory being used for each repeated action. Even selecting/deselecting or repeated renders seem to get the reported memory to over 6GB.... Any idea how to get an accurate report?

Thomas Wilmot commented

2015-08-03 03:25:51 +02:00

In #45629#325321, @sindra1961 wrote:
@ThomasWilmot:
Do you know this?
NVIDIA's Latest Windows 10 Drivers: Still Too Unstable For Primetime

That article is almost 5 months old now and refers to both laptop GPU's and experimental drivers, not desktop cards nor WHQL drivers :)

In relation to Martijn, I'm using an SLI bridge. When Blender is open, my cards use about 1730MB when not rendering, and when blender is closed, it uses about 1450. When I render, GPUz says I'm using -4191913MB, so, um... yeahhh, not real sure whats going on there :P

I tried changing my resolution to 1080p, and my cards only use about 1030MB without blender open, and when I replicate the bug, I can get to a higher RAM usage before the bug appears.

I don't really understand how GPU memory works, or how CUDA handles it, but is it possible that it's filling up the memory on one card, trying to move it to the other card and then running into a pointer issue, or something of the like?

> In #45629#325321, @sindra1961 wrote: > @ThomasWilmot: > Do you know this? > [NVIDIA's Latest Windows 10 Drivers: Still Too Unstable For Primetime ](http://www.dailytech.com/NVIDIAs+Latest+Windows+10+Drivers+Still+Too+Unstable+For+Primetime/article37267.htm) That article is almost 5 months old now and refers to both laptop GPU's and experimental drivers, not desktop cards nor WHQL drivers :) In relation to Martijn, I'm using an SLI bridge. When Blender is open, my cards use about 1730MB when not rendering, and when blender is closed, it uses about 1450. When I render, GPUz says I'm using -4191913MB, so, um... yeahhh, not real sure whats going on there :P I tried changing my resolution to 1080p, and my cards only use about 1030MB without blender open, and when I replicate the bug, I can get to a higher RAM usage before the bug appears. I don't really understand how GPU memory works, or how CUDA handles it, but is it possible that it's filling up the memory on one card, trying to move it to the other card and then running into a pointer issue, or something of the like?

Thomas Wilmot commented

2015-08-03 03:48:10 +02:00

In #45629#325356, @ThomasDinges wrote:
To clarify, does it work with Blender 2.74 on Windows 10? Or was that tested on a older Windows OS?

Setting to Low for now, likely a driver problem and nothing we can fix right now.

Can confirm that the error occurs on Blender 2.74 on Windows 10.

> In #45629#325356, @ThomasDinges wrote: > To clarify, does it work with Blender 2.74 *on* Windows 10? Or was that tested on a older Windows OS? > > Setting to Low for now, likely a driver problem and nothing we can fix right now. Can confirm that the error occurs on Blender 2.74 on Windows 10.

Martijn Berger commented

2015-08-03 09:08:02 +02:00

Your NVIDIA GPU always uses 6-16 megs of ram as the driver needs some.
Any screen you connect will eat ram ( more if you have desktop compositing / transparent windows etc enabled ) ( can be around 600 megs for 2 megapixel display easy )
When blender starts using cuda and loading a kernel ( can be anywhere between 300 and 1200 megs, depending on experimental features, tile size, passes)
The actual scene and textures ( can be how ever many you still have left )

Blender can and will only report the 4th kind of ram use. The other kinds we cannot report per category as the API to do detailed GPU ram usage is obviously only useful for pro's that buy $4000 quadro's

- Your NVIDIA GPU always uses 6-16 megs of ram as the driver needs some. - Any screen you connect will eat ram ( more if you have desktop compositing / transparent windows etc enabled ) ( can be around 600 megs for 2 megapixel display easy ) - When blender starts using cuda and loading a kernel ( can be anywhere between 300 and 1200 megs, depending on experimental features, tile size, passes) - The actual scene and textures ( can be how ever many you still have left ) Blender can and will only report the 4th kind of ram use. The other kinds we cannot report per category as the API to do detailed GPU ram usage is obviously only useful for pro's that buy $4000 quadro's

Sergey Sharybin commented

2015-08-03 11:00:38 +02:00

Changed status from 'Open' to: 'Archived'

Sergey Sharybin closed this issue

2015-08-03 11:00:38 +02:00

Sergey Sharybin commented

2015-08-03 11:00:38 +02:00

You're using Experimental kernel which is rather really memory hungry due to the kernel itself and the per-thread stack memory it needs.

With "regular" kernel i was able to render with gtx580 which has 3gig of ram here, but it was reallty bouncing on the memory limit threshold.

Proper solution for avoiding huge kernel/stack memory overhead is to go microkernels (similar to what AMD was doing with OpenCL), but wrapping that stuff up and making it working on CUDA will take a bit of a time.

So thanks for the report, but it's not really a bug, just optimization which we need to do in the kernel.

You're using Experimental kernel which is rather really memory hungry due to the kernel itself and the per-thread stack memory it needs. With "regular" kernel i was able to render with gtx580 which has 3gig of ram here, but it was reallty bouncing on the memory limit threshold. Proper solution for avoiding huge kernel/stack memory overhead is to go microkernels (similar to what AMD was doing with OpenCL), but wrapping that stuff up and making it working on CUDA will take a bit of a time. So thanks for the report, but it's not really a bug, just optimization which we need to do in the kernel.

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Cycles - Out of Memory Bug #45629