VRED 2021 GPU Raytracing Requirements

[sschultz1]

What is the minimum dual GPU requirement to run the new GPU Raytracing within VRED? Also Is there a different minimum requirement to run the GPU Raytracing and the GPU Denioser at the same time?

[Christian_Garimberti · Accepted solution]

Hi,

as Pascal wrote in the news:

”VRED GPU Raytracing works on NVIDIA GPUs of the Turing, Volta, Pascal, and Maxwell architectures. RTX GPUs are highly recommended and multiple GPUs will improve performance. For optimal performance with 2x GPUs, enabling SLI and using a NVLink-Bridge is recommended. For more than 2 GPUs, the settings are not relevant. 1x VRED Pro license supports 2x GPUs. Only homogeneous GPU setups are supported. Your NVIDIA Driver version must be 441.6 or newer.”

i don’t think the use of the denoiser change the minimum requirement.

Best

Chris

[michael_nikelsky · Accepted solution]

To add to that: Of course the denoiser works with the GPU raytracing as well. The denoiser also has the same miminum requirements since it is also based on the current Optix version.

[sschultz1]

Thank you for the info. I'm a student and currently have a single Nvidia RTX 2060. Which is not enough by itself to use the GPU Raytracing option. It runs out of memory. Or at least it did with my current scene. Do you know if running two RTX 2070 Supers through NVLink would be adequate to see an advantage over CPU Raytracing? Or would a Single RTX 2080 ti be a better route? I can only assume running two of the  RTX 2080 ti's should show a benefit over CPU Raytracing, but obviously as a student I'm trying to optimize my performance and cost to achieve it. So I figure I can either buy one RTX 2080ti and add an additional one later in the future, or buy two RTX 2070 Supers. Do you have any incite or recommendations?

Best

Scott

[michael_nikelsky · Accepted solution]

We currently don´t support memory expansion with nvlink so having more memory on a single GPU would be better. So a 2080TI would probably be better and depending on your CPU you should see quite a performance improvement from it. We did a couple of tests comparing a single Quadro RTX 8000, which is only slightly faster than the 2080TI, to a 40Core/80Thread CPU system and it was more than twice as fast.

[Anonymous]

Hi,

I'm also a student with an RTX 2060 card, I'm learning how to use VRED and I want to use the GPU raytracing but when I go to enable it there's only option is CPU raytracing, the GPU option isn't even showing up. My drivers are up to date and have enough memory, if anyone can help would be much appreciated. 

[michael_nikelsky]

You need to have VRED Pro, not VRED Design for GPU raytracing at the moment. 

Other than that it should work on a 2060, at least for small scenes due to the limited memory.

[Anonymous]

FROM WHERE CAN I DOWNLOAD VRED PRO STUDENT VERSION?

[Christian_Garimberti]

From Here you have access to everything

https://www.autodesk.com/education/edu-software/overview?sorting=featured&page=1

best

Chris

[keshirasawa]

I prepared a server with 8 RTX A6000s. SLI is enabled (4 pairs) and geometry and texture memory sharing is turned on in VRED's general rendering settings.

I started VRED Pro, loaded about 16GB of VRED scene data, and tested ray tracing with a GPU.

I found that all 8 GPUs occupy about 46GB of GPU memory, and when I add more geometry, VRED doesn't go down and the CPU memory goes up over a very long time.
But the problem is that I may have to wait an hour for the rendered image to come out.
Do you know why this phenomenon is so time consuming?

[michael_nikelsky]

If you are already using 46GByte of the available 48GByte chances are you run into a memory issue due to fragmentation or pretty much anything else if you add more Geometry. Did you check the Share Texture and Share Geometries option in the Preferences (Render Settings->General Settings Tab)? This can help reducing the memory needed per GPU (note that this option currently only works on Windows, we are working on supporting linux for this as well but it will take a while).

Note that the first launch after installing a new driver or VRED version will take quite a long time since all GPUs have to compile their shaders. I am not sure how parallel this is happening since it is done in the driver but it can take several minutes.

Also make sure you start GPU Raytracing or CPU Raytracing before loading the dataset. At the moment we don´t clear the Memory OpenGL uses so in case of GPU Raytracing this means your scene will exist twice on the GPU if you load it while in OpenGL mode.

Kind regards

Michael

[keshirasawa]

Dear Micheel san,

Thank you for your reply. 

Even if VRED Pro memory sharing and geometry sharing were turned on, all GPU memory consumption was the same. 

What version of NVIDIA driver for RTX A6000 and A40 this feature works fine? 

Thank you for your cooperation.

Kind Regards,

Keiji Shirasawa,

[michael_nikelsky]

Hi,

it works fine with 2022.0 and driver 466.11. But as I said: If you are already at 46GBytes it might just be that case that your scene is too large to work with GPU raytracing. Memory Sharing will not double the amount of usable memory you have, that is just marketing nonsense. Please verify whether it works or not with a smaller scene:

Turn memory sharing off, restart VRED, activate GPU raytracing, load your scene and under Help->GLInfo check how much memory is used.

Then Turn memory sharing on, restart VRED, activate GPU raytracing, load your scene and under Help->GLInfo check how much memory is used now. It should be less since the data is distributed between the two GPUs of an island. If it is not less memory sharing does not work for some reason, mostlikely some issue with the NVLink connection.

You can also start VRED with the -console parameter. When turning on GPU raytracing it should print something like this for all the devices in your system:
Peer to peer connection enabled between devices 0 and 1
Peer to peer connection enabled between devices 1 and 0
Devices in Island for device 1:
Device 0 Island Id: 0
Device 1 Island Id: 1
Devices in Island for device 0:
Device 0 Island Id: 0
Device 1 Island Id: 1

Kind regards

Michael

[keshirasawa]

Dear Michael san,

I saw the difference between turning VRED texture sharing and geometry sharing on and off.

When I loaded a VRED scene of about 10GB and turned it on, the GPU memory consumption was about 40GB, and when it was off, it consumed about 46GB.
However, it takes more than 40 minutes for the GPU ray tracing screen to start.
Thank you for confirming!

Kind Regards,

Keiji Shirasawa,

[michael_nikelsky]

40 minutes doesn´t sound right. This hints to something being very strange about the scene. How many geometries and nodes are in the scene? Throwing lots of small geometries at the GPU raytracer is the worst thing to do since it causes an extreme overhead. Also how is your main memory during the initialization process? A rule of thumb is that you need at least as much main memory as you have GPU memory, so effectively I would say 128GByte main memory should be the minimum available.

[keshirasawa]

The CPU memory of VRED immediately after loading is about 39GB.

I'm then switching to GPU raytracing mode, and since I'm doing it now, I'll probably report GPU consumption and CPU memory consumption after 30-40 minutes. Later.

Kind Regards,

Keiji Shirasawa,

[michael_nikelsky]

Take a look at the memory consumption in the task manager while it is loading. If it reaches the maximum it starts swapping to disk which probably explains why it is so slow.

I just found that there is a bug in the 2022 and newer releases it seems, I am investigating it now. I you are able to you could test the 2021.3 version to see if it behaves better.

[keshirasawa]

I'm still waiting because the GPU ray tracing screen isn't displayed yet.
In version 2021.3, A6000 and A40 could not be displayed in a cluster on another GPU server. Since it was a pitch black screen, I am using the latest 2022.
Did you know this 2021.3 screen sync bug?

Kind Regards,

Keiji Shirasawa,

[keshirasawa]

The GPU raytracing is displayed.

At this time, the memory consumption of VRED of the task manager is about 94GB.

For GPU memory sharing, both texture and geometry are turned on, and GL information from HELP is about 39.9GB.
Also, this about 10GB is intentionally made huge data for customers to benchmark.

There are so many nodes,  5000 triangular polygon Spheres ｘ950 and 25 Automotive Genesis.

Kind Regards,

Keiji Shirasawa,

[michael_nikelsky]

As I said, many nodes will hurt the GPU Raytracer a lot. It needs to build each acceleration structure sequentially so the time to start GPU raytracing will increase at least linearly with the number of nodes (probably even more due some lookups). There is also some excessive memory consumption at the moment I am investigating (probably caused by the new lightmap feature).