Complex Blender scene crashes with bluescreen

[kode54]

When attempting to render this Blender scene using Blender 3.4.1 using oneAPI GPU Compute Cycles renderer, with the v3959 drivers:

https://f.losno.co/Rivet_blender_model.7z - 216.4 MB (reposted because original requires free registration to download)

It uses up to 20GB of system RAM on its own to build the BVH tree, then settles down to about 4-6GB. Then it hits "Importance map", and bluescreens the machine. The crash the first few times I tried it was an exception being thrown in Ntfs.sys, but then later, it was an exception in the DirectX graphics kernel.

This is with an A750 card, so maybe I am running out of VRAM, but that shouldn't crash the entire machine.

Edit: I have tested this again under Linux, using intel-compute-runtime-bin 22.35.24055. Unlike Windows, it actually errors out non-fatally. It ran out of VRAM. I guess if I want to build this model, I need an A770 with 16GB of VRAM. But at least the Windows drivers shouldn't throw a bluescreen on an out of memory condition.

[Stefan3D]

Nice sample,

it even overcharges a high-end NVIDIA Ampere GPU 😓

Try Blender via blender_debug_gpu.cmd

I got a gazillion entries in the log reading:

INFO (gpu.debug): Notification : Texture state usage warning: The texture object (0) bound to texture image unit 0 does not have a defined base level and cannot be used for texture mapping.

And a bunch of other warnings, eg

INFO (gpu.debug): Notification : Program/shader state usage warning: A fragment program/shader is required to correctly render to an integer framebuffer.

[Robert_Q_Intel]

Hello kode54,

Thank you for posting on the Intel®️ Arc communities.

We understand the issue that you are experiencing with your Intel® Arc™ A750 Graphics.

• To provide you with the support we would like to gather the following information:

• When did the issue start?

• Please take screen captures of the usage of the memory, and post them in your reply

•  Download and run Intel® System Support Utility for Windows* make sure to check the box "Everything"and save the resulting .txt file you can attach it to your reply.Link for Intel® System Support Utility for Windows* https://www.intel.com/content/www/us/en/download/18377/intel-system-support-utility-for-windows.html

Best regards,  

Robert Q. 

Intel Customer Support Technician

[kode54]

I cannot capture those screen shots, for several reasons.

1) The memory usage happens all at once the instant it reaches a certain point in the Blender render process, at which point it immediately clobbers the system memory and crashes the system.

2) I am no longer running Windows 11. The system install in question has been backed up to a Clonezilla image, and replaced with Arch Linux.

3) From Arch Linux, I attempted the same exact render. It used slightly less system memory building the BVH trees, but again failed to upload them to the GPU when it reached the "Importance map" stage of the render process. This time, however, it emitted a oneAPI error about insufficient VRAM, and that it needed approximately 7.6GB of VRAM all to itself, on top of whatever the system was using. This would be too much, considering my average GNOME session is using approximately 1.1GB of VRAM at the steady state of my usual applications running, and Windows was using approximately 2.3GB of VRAM with a similar set of apps running.

This error started within a day or two of installing a new Windows 11 Pro environment, and deciding to test some Blender scenes on it to run the hardware through its paces. I wanted to see if the GPU could beat my CPU at rendering that scene, which would have taken about an hour to run the 1000 samples it's configured for.

I suppose I could capture a screen shot of Blender on Linux showing the OOM error, but is that going to be terribly useful? Other than downloading 200MB of data from my CDN hosting, and running a render process on it for approximately 5 minutes, it should render just fine on an A770 with 16GB of VRAM, while having possible problems on an A750. If it doesn't cause an outright system crash, perhaps the issue is fixed now?

[Carlos_L_Intel]

Hi @kode54

Thank you for your post. I completely understand. Can you provide us with more details of your system so we can help you better:

• Model of your motherboard and BIOS version.
• When in Windows, was this a clean installation of the OS? 
• Were there any cards previous install? 
• Is Rezaisable Bar enabled? 

Best regards,  

Carlos L.  

Intel Customer Support Technician

[kode54]

• Model of your motherboard and BIOS version.

MSI B450 Tomahawk, running 7C02v1I, the latest beta BIOS.

• When in Windows, was this a clean installation of the OS? 

Yes, a clean install. Windows was installed freshly after aquiring the new A750 video card.

• Were there any cards previous install? 

No previous cards associated with this Windows 11 install.

• Is Rezaisable Bar enabled? 

Resizable BAR is supported by this BIOS revision, and is enabled. I also verified that it worked with my previous video card in a previous install of Linux. Yes, I know that AMD's drivers only allow it with Ryzen 5000 series and newer, but it works with my 2700.

[Robert_Q_Intel]

Hello kode54,

Thank you for posting,

We will start reviewing this issue internally, in the meantime, we would like to provide you with the following recommendations:

recently a new version of the Intel® Arc™ & Iris® Xe Graphics - WHQL - Windows*, was released the version number is 31.0.101.4032, please install this new version and test the game in order to confirm if there is any signs of improvement under this new version

To help you download the Intel® Arc™ Graphics - WHQL - Windows* driver please visit the following link:

https://www.intel.com/content/www/us/en/download/726609/intel-arc-iris-xe-graphics-whql-windows.html

To find a guide that will help you while completing a clean installation of this driver:

https://www.intel.com/content/www/us/en/download/726609/intel-arc-iris-xe-graphics-whql-windows.html

Remember to Download and run Intel® System Support Utility for Windows* make sure to check the box "Everything" and save the resulting .txt file you can attach it to your reply. Link for Intel® System Support Utility for Windows* https://www.intel.com/content/www/us/en/download/18377/intel-system-support-utility-for-windows.html

this report will provide us with general information about your device in order to enhance the research about the issue that you are experiencing.

Once you have completed these steps please share the outcome with us, we will be waiting for your response in case we don't receive it we will follow up in 3 business days.

Best regards,  

Robert Q. 

Intel Customer Support Technician

[kode54]

I couldn't reply to this topic with my Microsoft Edge setup, but reply works just fine in Mozilla Firefox. Sorry for creating a duplicate topic.

This scene appears to push GPU VRAM up to 14GB, but renders successfully on an Arc A770, now that I've got one of those instead of the A750 I returned. It would be nice if the drivers more gracefully handled the Out Of Memory situation that Blender / oneAPI causes, instead of having an internal driver exception, or worse, possibly clobbering memory.

[Robert_Q_Intel]

Hi @kode54

Thank you very much for the feedback we really appreciate it.

Now that you are using an Intel® Arc™ A770 beside the high usage of VRAM memory with this specific scene are you experiencing any particular issues with the VRAM usage in other applications/tasks?

Best regards,  

Robert Q. 

Intel Customer Support Technician

[Robert_Q_Intel]

Hi @kode54,

Were you able to check the previous post?  

Let us know if have more details about this behavior.

Best regards,  

Robert Q. 

Intel Customer Support Technician

[Robert_Q_Intel]

Hi @kode54,

We have not heard back from you, so we will close this thread. If you need any additional information, please submit a new question as this thread will no longer be monitored.

Best regards,  

Robert Q. 

Intel Customer Support Technician

[Jean_Intel]

Hello @kode54,

As you mentioned that your CPU is an AMD 2700 CPU, we would like to inform you that for optimal performance, we recommend the 3000 series or newer as the requirements indicate.

Also, since the troubleshooting is completed and the issue occurs with others GPUs, it indicates there is an Operating System/App conflict. We recommend you report this problem to the app developer for further recommendations.

Best regards,

Jean O. 

Intel Customer Support Technician

[kode54]

As I've already stated, the problem went away when I returned the A750 and purchased a more expensive A770 with 16GB of VRAM, and re-ran the render to find that Blender wanted to occupy 14GB of VRAM.

Not necessarily a problem caused by having below recommended CPU and motherboard. You would also have me replace my motherboard with either an X570 or B550 as well, since I have a B450 board currently, which is an unsupported 400 series. I highly doubt the board or processor have anything to do with memory exhaustion in a special circumstance causing a driver crash. But I guess just close this as solved but not solved until I find something that exhausts 16GB completely and buy the supported board and processor configuration. By that pricing, I may as well just skip buying any future AM4 parts and just buy a shiny new 13600k instead. And all new RAM to boot.