concerning benchmark results on my GPU vs CPU - mike pan benchmark file

Support Technical Support
1

Hey, I’m a bit concerned about what is going on with my GPU. I have a Nvidea 550TI.

I’ve been noticing that my renders seem to be unusually slow on GPU … My GPU has always been faster when rendering in cycles than my CPU provided the scene can fit within my 1 gb GPU memory.

So I just did the mike pan benchmark and my results

I running windows 7 ,64 bit, blender2.71 official, this is my CPU

Intel Core i7 2600 @ 3.40GHz
Sandy Bridge 32nm Technology

My concern is my GPU results are nearly identical to my CPU results. 2:01 GPU 2:16 CPU

2 things that have been done recently are…

-I did upgrade GPU drivers a few weeks ago
-I made a time out registry change to correct the problem with GPU crashing when viewport rendering due to time out errors.
-I’m not getting any unusually high temp readings for GPU under load

  1. Is there any other tools available to check the “health” of your GPU?
  2. Any suggestions on what else I should do
2

Try really cranking up the samples and issue another render so the GPU has some time to get up to full load. I use the MSI Afterburner application to monitor and control my GPU fan.

I read about the registry timeout “fix” and got mixed reviews as to whether or not it really works or just throws another variable into the equation. Try resetting that value to default.

Try altering the tile size. I would think somewhere in the range of 96-384.

3

There is a much better performance on CPU with Blender 2.71 compare to 2.70, GPU slows down.
2.72 is the same process.
I think you know but you need big tiles on GPU, the Mike Pan file is really bad for GPU with default settings.

Cheers, mib

4

This is unfortunately true!

5

@atom I have to say, since making the change atom, I have not had one single timeout crash, let me ask you something. I’m not opposed to rolling back, but since I made the change I have done the following…

  • updated video drivers
  • There has been numerous windows updates after making the change
  • I manually went in and changed some settings to change the timeout stuff, but prior to doing that I did 2 things… (1) made a system restore (2) I also made some kind of registry backup but I cant remember if it was only on the registry area I was changing. I don’t think it was a full registry backup.

Do you think system restore is the better way to go if I want to roll back?

-before doing any rollbacks, I want to try your other suggestions first regarding the increased samples and changing the tile size… I think I just left it at what ever the default was when you opened the file, and only changed the cpu / gpu setting and rendered.

@mib yea, I definitely saw an improvement in CPU speed when moving to 2.71. You once had my same GPU if I remember correctly.
But did the GPU slow down that much? There is no volumtrics, sss, in the mike pan benchmark, so are you saying the GPU slowed down that much just on regular scenes from 2.71 to 2.72?

@ everybody

  1. All I know is my GPU has been way way faster than CPU up until now. Does anyone think the change in my drivers could be an issue?

  2. Is there a better benchmark test I should try besides the Mike Pan one?

6

I confirm GPU slow down for cycles in 2.72 compared to 2.71 and 2.70

i5-4670, GTX 750Ti OC (MSI) without overclock, 8GB RAM, win 8.1pro

Mike pan’s BMW (256x256 tiles):
2.71 = 1:00.21
2.72RC = 1:10.58 (+17% render time)

Cornell box 2.71 benchmark (stock file):
2.71 = 7:16.91
2.72RC = 9:04.07 (+25% render time)

These times looks like those I obtained testing DingTo’s 2.71 build compiled with CUDA 6.5 (BMW: +12.5%, Cornell: +17%), could there be any link ?

Anyone experimenting such slowdowns on 750/750Ti ?

EDIT: Mike pan’s bench is done at 256x256 tiles, not stock

7

I’m seeing slowdowns as well on a 750 ti. Using the Mike Pan BMW as an example, I’ve gone from 1:16 with 2.71 to 1:29 with 2.72 RC (both builds downloaded from blender.org)

8

Meaning +17% render time as well…
Thank for sharing, I’ll ask other people using 750 and 750ti

9

@mib yea, I definitely saw an improvement in CPU speed when moving to 2.71. You once had my same GPU if I remember correctly.
But did the GPU slow down that much? There is no volumtrics, sss, in the mike pan benchmark, so are you saying the GPU slowed down that much just on regular scenes from 2.71 to 2.72?

Yes and yes, each feature add to cycles GPU slow down the engine if you use it or not, I fear.
I use the GTX 550Ti for some time but it was still to slow compare to my i5 Ivy so I add the 560Ti 448 and use the 550Ti for display. This was a good duo for long time.
Btw. SSS on GPU use a different kernel (Experimental) because it is again 10% slower. :expressionless:

Cheers, mib
EDIT: Argh, missed my 2,222 post. :stuck_out_tongue:

10

@mib… you mentioned you still use that GTX550ti for display purposes… I have no idea if this would be possible for you to do… if it involves too much work than please disregard…

But is there anyway you can run that Mike pan benchmark using default settings on your 550ti? I would just be really curious just to see how long your GPU takes to render it?

11

Hi, I am sorry, my English is not good enough.
I work a long time with 550 and 560 but kick the 550 and buy a GTX 760 4 GB
so I cant test it with GTX 550Ti.
You can try my benchmark file but nobody with GTX 550Ti report in the whole thread.

http://www.blenderartists.org/forum/showthread.php?342000-The-new-Cyles-GPU-2-71-Benchmark

Cheers, mib