Cycles GPU performance regression on Nvidia's Maxwell cards #47273

New Issue

Antonio Buch · 2016-01-29T10:22:53+01:00

Antonio Buch commented

2016-01-29 10:22:53 +01:00

System Information
Windows 10
GeForce GTX TITAN Black
GeForce GTX 970
GeForce GTX TITAN X

Blender Version
Broken: 2.76b f337fea
Works: 2,73a bbf09d9

Short description of error
It looks like there is a regression in the performance of Nvidia's Maxwell generation cards on every new Blender version. At least with the GTX 970 and TITAN X which are the cards we have.

We've noticed this regresion since version 2.73a, when we got our TITAN X in April 2015. At that time, the TITAN X performed better than TITAN Black but in the current version (2.76b) the TITAN Black performs way better than TITAN X. The TITAN X is performing even worse than the GTX 970.

Steps for others to reproduce the error
Render the same scene with Blender 2.73a and 2.76b. You'll find that Kepler cards (like TITAN Black) performs better on 2.76b while Maxwell ones performs better on 2.73a than newer versions.

Here are some test results of 3 computers with GTX 970, TITAN Black and TITAN X at 2 different tile sizes and two different Drivers for TITAN X. The scene rendered was the BMW benchmark https://www.dropbox.com/s/6eg8a4l4lqig7sh/BMW27.blend?dl=0

TITAN Black (Tile 240x270)

2.76b – 01:03.25
2.73a – 01:05.35

TITAN Black (Tile 480x540)

2.76b – 00:59.83
2.73a – 01:01.37

GTX 970 (Tile 240x270)

2.76b – 01:28.26
2.73a – 01:18.94

GTX 970 (Tile 480x270)

2.76b – 01:25.16
2.73a – 01:17.16

GTX TITAN X (Tile 240x270) Nvidia Driver V. 361.43 (the latest)

2.76b – 01:31.23
2.73a – 00:59.24

GTX TITAN X (Tile 480x540) Nvidia Driver V. 361.43 (the latest)

2.76b – 01:21.74
2.73a – 00:50.64

GTX TITAN X (Tile 240x270) with Nvidia Driver V. 352.84 (the newer at April '15)

2.76b – 01:35.28
2.73a – 01:01.40

GTX TITAN X (Tile 480x540) Nvidia Driver V. 352.84 (the newer at April '15)

2.76b – 01:26.96
2.73a – 00:52.35

**System Information** Windows 10 GeForce GTX TITAN Black GeForce GTX 970 GeForce GTX TITAN X **Blender Version** Broken: 2.76b f337fea Works: 2,73a bbf09d9 **Short description of error** It looks like there is a regression in the performance of Nvidia's Maxwell generation cards on every new Blender version. At least with the GTX 970 and TITAN X which are the cards we have. We've noticed this regresion since version 2.73a, when we got our TITAN X in April 2015. At that time, the TITAN X performed better than TITAN Black but in the current version (2.76b) the TITAN Black performs way better than TITAN X. The TITAN X is performing even worse than the GTX 970. **Steps for others to reproduce the error** Render the same scene with Blender 2.73a and 2.76b. You'll find that Kepler cards (like TITAN Black) performs better on 2.76b while Maxwell ones performs better on 2.73a than newer versions. Here are some test results of 3 computers with GTX 970, TITAN Black and TITAN X at 2 different tile sizes and two different Drivers for TITAN X. The scene rendered was the BMW benchmark https://www.dropbox.com/s/6eg8a4l4lqig7sh/BMW27.blend?dl=0 **TITAN Black (Tile 240x270)** 2.76b – 01:03.25 2.73a – 01:05.35 **TITAN Black (Tile 480x540)** 2.76b – 00:59.83 2.73a – 01:01.37 **GTX 970 (Tile 240x270)** 2.76b – 01:28.26 2.73a – 01:18.94 **GTX 970 (Tile 480x270)** 2.76b – 01:25.16 2.73a – 01:17.16 **GTX TITAN X (Tile 240x270) Nvidia Driver V. 361.43 (the latest)** 2.76b – 01:31.23 2.73a – 00:59.24 **GTX TITAN X (Tile 480x540) Nvidia Driver V. 361.43 (the latest)** 2.76b – 01:21.74 2.73a – 00:50.64 **GTX TITAN X (Tile 240x270) with Nvidia Driver V. 352.84 (the newer at April '15)** 2.76b – 01:35.28 2.73a – 01:01.40 **GTX TITAN X (Tile 480x540) Nvidia Driver V. 352.84 (the newer at April '15)** 2.76b – 01:26.96 2.73a – 00:52.35

Antonio Buch commented

2016-01-29 10:22:53 +01:00

Changed status to: 'Open'

Antonio Buch commented

2016-01-29 10:22:54 +01:00

Added subscriber: @antoniobuch1

Antonio Buch commented

2016-01-29 10:33:16 +01:00

I'm not sure how Priority works so I set at its "Needs triage" default. But as you can see, the "bug" is tested and I really want this task as High Priority. It is slowing down our studio production and our investment in the TITAN X is pointless now.

Thomas Dinges commented

2016-01-29 13:37:58 +01:00

Added subscriber: @ThomasDinges

Thomas Dinges commented

2016-01-29 13:37:58 +01:00

Changed status from 'Open' to: 'Archived'

Thomas Dinges closed this issue

2016-01-29 13:37:58 +01:00

Thomas Dinges self-assigned this 2016-01-29 13:37:58 +01:00

Thomas Dinges commented

2016-01-29 13:37:58 +01:00

Hi,
I understand your concern here, and we are dedicated to make Cycles as fast as possible. Unfortunately, new features and code changes can always result in performance changes and drops, that is very difficult to avoid.

We plan on porting the split kernel architecture to CUDA as well, which should improve performance quite a bit again.

We consider this an ongoing project, not something that can be solved in a day.

Hi, I understand your concern here, and we are dedicated to make Cycles as fast as possible. Unfortunately, new features and code changes can always result in performance changes and drops, that is very difficult to avoid. We plan on porting the split kernel architecture to CUDA as well, which should improve performance quite a bit again. We consider this an ongoing project, not something that can be solved in a day.

Antonio Buch commented

2016-01-29 13:50:51 +01:00

Hi Thomas, glad to meet you. Of course it cannot be solved fast, though I have no idea of programming I'm sure it is a tough work developing a render engine such as Cycles. I'm reporting this just in case you don't tested with Maxwell GPU's and you don't know this issue.

So did you know about this behavior? Are there plans to solve this? I don't want to be too nosy, it's just for planning future purchases. Thanks!

Hi Thomas, glad to meet you. Of course it cannot be solved fast, though I have no idea of programming I'm sure it is a tough work developing a render engine such as Cycles. I'm reporting this just in case you don't tested with Maxwell GPU's and you don't know this issue. So did you know about this behavior? Are there plans to solve this? I don't want to be too nosy, it's just for planning future purchases. Thanks!

Thomas Dinges commented

2016-01-29 13:55:23 +01:00

Added subscriber: @Sergey

Thomas Dinges commented

2016-01-29 13:55:23 +01:00

I am not sure we have Titan X cards at the Blender Institute (@Sergey will know), but I cannot remember a specific change, that could cause such a slowdown. Can you maybe test with a build from the Buildbot (https://builder.blender.org/download/) and see how that one performs on the Titan X? Is it worse than 2.76b or better?

We will work on the Split kernel architecture soon, that will improve performance on CUDA GPUs hopefully, but how much this will help is yet to be seen.

Developer note:

  Titan Black = sm_35
  Titan X = sm_52

I am not sure we have Titan X cards at the Blender Institute (@Sergey will know), but I cannot remember a specific change, that could cause such a slowdown. Can you maybe test with a build from the Buildbot (https://builder.blender.org/download/) and see how that one performs on the Titan X? Is it worse than 2.76b or better? We will work on the Split kernel architecture soon, that will improve performance on CUDA GPUs hopefully, but how much this will help is yet to be seen. Developer note: ``` Titan Black = sm_35 Titan X = sm_52

Antonio Buch commented

2016-01-29 14:13:21 +01:00

Just checked. The result in the last build (e2161ca) is exactly the same.

Just checked. The result in the last build (e2161ca) is exactly the same.

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 GPU performance regression on Nvidia's Maxwell cards #47273