Inefficient use of GPU? #91791

New Issue

Kepa · 2021-09-28T20:23:09+02:00

Kepa commented

2021-09-28 20:23:09 +02:00

System Information
Operating system: Windows-10-10.0.19041-SP0 64 Bits
Graphics card: NVIDIA GeForce RTX 2080 Ti/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 472.12

Blender Version
Broken: version: 3.0.0 Alpha, branch: master, commit date: 2021-09-27 18:29, hash: c53ffda8a4
Worked: (newest version of Blender that worked as expected)

Short description of error
The total use of the GPU when rendering has dropped by around 10-15% according to different monitoring with different software that I have been doing between the latest LTS version (2.93.4) and the ALPHA (3.0.0).
The scene is the same in both cases.

The team is made up of 4 Nvidia 2080ti GPUs and the tests have been carried out both one by one, two, three or all 4, the result always being the same.
I am attaching two screenshots and I hope it will be useful.

Exact steps for others to reproduce the error
Simply compare the use of the GPUs with some monitoring software, while the render is running. In my case TechPowerUp GPU-Z, in its latest version.
2.93.4

3.0.0 Alpha

**System Information** Operating system: Windows-10-10.0.19041-SP0 64 Bits Graphics card: NVIDIA GeForce RTX 2080 Ti/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 472.12 **Blender Version** Broken: version: 3.0.0 Alpha, branch: master, commit date: 2021-09-27 18:29, hash: `c53ffda8a4` Worked: (newest version of Blender that worked as expected) **Short description of error** The total use of the GPU when rendering has dropped by around 10-15% according to different monitoring with different software that I have been doing between the latest LTS version (2.93.4) and the ALPHA (3.0.0). The scene is the same in both cases. The team is made up of 4 Nvidia 2080ti GPUs and the tests have been carried out both one by one, two, three or all 4, the result always being the same. I am attaching two screenshots and I hope it will be useful. **Exact steps for others to reproduce the error** Simply compare the use of the GPUs with some monitoring software, while the render is running. In my case TechPowerUp GPU-Z, in its latest version. 2.93.4 ![2.93.4_version.jpg](https://archive.blender.org/developer/F10654862/2.93.4_version.jpg) 3.0.0 Alpha ![3.0.0 Alpha_version.jpg](https://archive.blender.org/developer/F10654868/3.0.0_Alpha_version.jpg)

Kepa commented

2021-09-28 20:23:09 +02:00

Added subscriber: @Kepa-1

Jesse Yurkovich commented

2021-09-28 20:28:38 +02:00

Added subscriber: @deadpin

Jesse Yurkovich commented

2021-09-28 20:28:38 +02:00

There is still ongoing work in this area as detailed within #89833 (see last few comments).

Kepa commented

2021-09-28 20:34:52 +02:00

Hi Jesse,

Thanks for responding so quickly, I have read what you tell me but it is not clear to me if the error is due to the distribution of tasks when using multi-GPU or if this also happens in version 3.0.0 Alpha, with only one GPU.

The tests that I have carried out have been with a single GPU and with several and the result is the same, I don't know if I understand that what you tell me is only related to multi-GPUs.

Thank you.

Hi Jesse, Thanks for responding so quickly, I have read what you tell me but it is not clear to me if the error is due to the distribution of tasks when using multi-GPU or if this also happens in version 3.0.0 Alpha, with only one GPU. The tests that I have carried out have been with a single GPU and with several and the result is the same, I don't know if I understand that what you tell me is only related to multi-GPUs. Thank you.

Jesse Yurkovich commented

2021-09-28 20:42:57 +02:00

The mentioned bug is for multiple devices. If you are observing a rendering time regression compared to previous versions with just 1 device, please make your bug very clear it's just for 1 GPU and attach an example .blend file (as per the bug filing template and guidelines) so that it could be investigated.

Kepa commented

2021-09-28 20:59:13 +02:00

Thanks Jesse,

In the description of the problem I indicate that the result of the test is using a single GPU, two GPUs, three GPUs and the four GPUs, that is, the test result is the same with any number of GPUs, but if you need to report another message again where I only indicate that it is with one, tell me and I report it if it is better.

This happens with any scene, I don't know if it is relevant that I send you a scene because it happens even with a simple box.
Tell me if you need me to do something else to help and I will gladly help you.

Thanks Jesse, In the description of the problem I indicate that the result of the test is using a single GPU, two GPUs, three GPUs and the four GPUs, that is, the test result is the same with any number of GPUs, but if you need to report another message again where I only indicate that it is with one, tell me and I report it if it is better. This happens with any scene, I don't know if it is relevant that I send you a scene because it happens even with a simple box. Tell me if you need me to do something else to help and I will gladly help you.

Brecht Van Lommel commented

2021-09-28 21:04:02 +02:00

Added subscriber: @brecht

Brecht Van Lommel commented

2021-09-28 21:04:02 +02:00

GPU load being lower than 100% is not necessarily indicative of a bug. It could be that there is something to be optimized, but I would actually expect it to be lower with the cycles-x changes. What we would consider a bug is if there is a significant performance regression in render time or interactivity.

Previously we would schedule one big kernel for all paths, keeping all GPU cores occupied. However, due to divergence paths might terminate and keep cores occupied while not doing anything useful. Now when paths are terminated, we no longer schedule any work for them, and if few active paths remain that means not reaching 100% GPU load. However the effective occupancy might still be higher than before, just not visible in these kinds of statistics.

GPU load being lower than 100% is not necessarily indicative of a bug. It could be that there is something to be optimized, but I would actually expect it to be lower with the cycles-x changes. What we would consider a bug is if there is a significant performance regression in render time or interactivity. Previously we would schedule one big kernel for all paths, keeping all GPU cores occupied. However, due to divergence paths might terminate and keep cores occupied while not doing anything useful. Now when paths are terminated, we no longer schedule any work for them, and if few active paths remain that means not reaching 100% GPU load. However the effective occupancy might still be higher than before, just not visible in these kinds of statistics.

Kepa commented

2021-09-28 21:20:11 +02:00

Thanks Brecht,

I understand what you are telling me, I have decided to report it, because in identical conditions it seemed strange to me, that in the same scene the load was lower, as I showed you in the captures.

Thanks Brecht, I understand what you are telling me, I have decided to report it, because in identical conditions it seemed strange to me, that in the same scene the load was lower, as I showed you in the captures.

Richard Antalik commented

2021-09-29 17:31:57 +02:00

Added subscriber: @iss

Richard Antalik commented

2021-09-29 17:31:57 +02:00

Changed status from 'Needs Triage' to: 'Needs User Info'

Richard Antalik commented

2021-09-29 17:31:57 +02:00

@Kepa-1 do you see render time being worse in 3.0 alpha?

Kepa commented

2021-09-29 22:15:02 +02:00

Added subscriber: @richard-27

Kepa commented

2021-09-29 22:15:02 +02:00

Time is better in version 3.0.0 Alpha, but really what I expose would not be significant, I will not apply for what I say, because Cycles in version 3.0 Alpha even if it was 60-70% and it would be faster than in The 2.93.4, the difference that I notice is, that at least visually in the load of GPU is not done, the same load, but as Brecht says, not necessarily that has to mean that it is a bug, what I do not know It is if we even win more if that 10-15% more load that could be done on the GPU would still give us more performance.

Richard Antalik commented

2021-09-30 14:52:57 +02:00

Changed status from 'Needs User Info' to: 'Archived'

Richard Antalik closed this issue

2021-09-30 14:52:57 +02:00

Richard Antalik commented

2021-09-30 14:52:57 +02:00

Thanks for clarification. Since there is no performance regression I will close the report. As @brecht said, it is possible that there may be some room for improvement but from information provided here it is hard to tell.

For more information on why this isn't considered a bug, visit: https://wiki.blender.org/wiki/Reference/Not_a_bug

Thanks for clarification. Since there is no performance regression I will close the report. As @brecht said, it is possible that there may be some room for improvement but from information provided here it is hard to tell. For more information on why this isn't considered a bug, visit: https://wiki.blender.org/wiki/Reference/Not_a_bug

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

Inefficient use of GPU? #91791