Graphic card testing for Blender [Updated]

Some might have noticed by now I am on a crucade against poor OpenGL performance and to determine what card could be considered “best” for Blender :wink:

This endevour started because I wanted to find out if you can run a Radeon and GeForce in one System and still use CUDA/OpenCL. You´ll read it in the final report.
Short: Yes, it works flawlessly (besides color management) and I tested it with Octane, SLG and Lux.

Proof? Sure

I am still testing OpenGL performance and in the end I´ll hand out a nice PDF report, but I hit a dead end, especially with Radeon cards and my knowledge about their settings, not to mention I only got one.

However, with my testscene (download below) I get current results (exclusive preview for my report): GTX 470: 76fps (FW270, W7x64, [email protected], 1920x1200)
GTX 470: 51fps (FW268, W7x64, [email protected], 1920x1200)
GF 9500GT: 67fps (FW270, W7x64, [email protected], 1600x1200)
Quadro FX 1800: 42fps (FW270, W7x64, Ci7-950, 1920x1080)
HD 5850: 41fps (C10.4, W7x64, [email protected], 1920x1200)
GTS 250: 36fps (FW268, W7x64, Phenom2 X4-940, 1920x1080)
HD 5850: 35fps (C10.4, W7x64, Phenom2 X4-940, 1920x1080)
HD 5850: 22fps (C10.4, W7x32, K6-2/5000+, 1920x1200)
HD 5850: 9fps (C10.4, Ubuntu 10.10 x32, K6-2/5000+, 1920x1200)
You all can imagine my surprise that…
…the “crippled” Fermi is the fastest - and faster with the new ForceWare
…the cheapass 9500GT is incredibly fast. (60 Euro, passive cooled) I didn´t want to test it at first because I thought it´ll be substandard.
…the HD5850 performs so poor - which surprises me.

This first result basically makes fun of me, many know I am one of the advocates to spread the word on Fermi´s bad OpenGL performance, but it seems so far Fermi leads, but compared with the 9500GT it´s an ridiculous result.

Thats where I need the communities help!

First of all, the testing methodology:
You can either mail me your results to [email protected] if you don´t want to post, or, obviously post in this thread.

  • Be sure to turn OFF! vsync, else it will cap at your screens refresh rate.
  • Be sure to have VBO ON!
  • Use the highest resolution possible or your screen native resolution.
  • Use the Official Blender Builds, if this goes on longer, use v2.57b it´s what I started with.
  • Try to use WHQL/stable drivers

If you vary from those settings because you know your cards perform better with other settings, please note it in your report. If you don´t know how to do one or any of the above, please don´t participate :wink:

The scene is set to cap at 120 FPS which I consider plenty for 370k triangles. If you got the feeling there´s more in your card, subdivide the sphere once more, be sure to APPLY the modifier and run it again. Note the change in the form below as well and be sure the screenshot shows the amount of faces. In this case as it is an Icosphere faces = triangles.

What I need is (you can quote this):

Graphic card type:
Graphics driver version:
Number of graphic cards in system:
CPU type:
Mainboard type:
Operating System:
System Memory:
Screen Resolution:
Notes:
And a screenshot to have “proof” , not just the fps.

I´d prefere to have screenshots like this one, .jpg(100%) or .png else the font gets hard to read like below:


Just wide enought to see the amount of faces (forum cuts the image off, rightclick/view image or use the horizontal scrollbar at the bottom of the post) in the scene and high enough to see the fps for the PDF.

I need the memory amount, amount of cards and mainboard type to determine if any issues (low memory, PCIe slowdown due to several cards, bad chipset) are responsible for single rather odd (fast or slow) results

And I know it isn´t exactly scientific to determine the “best card” this way, if you want feel free to donate me cards to test them all in one refrence system :slight_smile: , but it should be good enough to determine where which card is ranked approximately and help the community out.

I especially want to find out if the 5850 idd has poor OGL performance too, or if it is my fault. I tried it in 3 systems though. And it seems very whatever else it is dependent. The weaker the system the more impact on the performance, but once you see the scene you´d notice there´s not much to do for the cpu.
It is a simple Icosphere with applied subdivision, doublesided=on and a camera flying around it endlessly. It has GLSL on and textured solid on. Sometimes it is not opening with camera view, so press num0 to do so if not. Start it with Alt+A
And if you turn off textured solid, at least my GTX470 performs much worse (-30fps)
That´s why I said “not very scientific”, we would have to benchmark solid view, solid view GLSL, solid textured, textured view, wireframe… but see the “Goal” section.
http://gallery.arexma.net/ba/Radeon_…estscene.blend

Now for the cards I need/want/miss (only a few :smiley: )

Nvidia Quadro 2000 (this one I need badly to compare against Fermi GeForce)
Nvidia Quadro 2000+ (any Fermi based Quadro - no Quadro FX)
Nvidia GeForce 285
Nvidia GeForce 260
Nvidia GeForce 460
Nvidia GeForce 560
Nvidia GeForce 480
Nvidia GeForce 570
Nvidia GeForce 580
Radeon HD 5850
Radeon HD 5870
Radeon HD 6850
Radeon HD 6870
Radeon HD 6950
Radeon HD 6970
Radeon HD 4850
Radeon HD 4870

If anyone want´s to supply “budget” or low-end cards I´ll take those too, but I don´t really need them because it´s more likely people will upgrade from those cards.

The goal:
Not sure. Satisfy my personal interest, find out if Fermi is really that bad, prevent the weekly “what´s the best graphics card for blender” - thread.

The long term goal however if it works out, maybe create a nice .py testing scene, running through various display modes, CPU performance and create something like the old blender benchmarking homepage, which sadly is discontinued.
I think it´s the perfect method to provide the community with information on hardware upgrades particular for blender and single out components that ain´t good for it.

Some trivia:
I tried one of my production plant prototype models (3.7 million polys) in Blender and Max on the GTX470 to compare viewport performance:
3dsmax 2011 x64 OpenGL: 0-1 FPS
3dsmax 2011 x64 DX10: 4 FPS
Blender 2.57b x64 OpenGL: 9 FPS

I just noticed something, which might render all this testing useless -.-
facepalm seems I tripped over my own eagerness.

My scene with the GTX470:
GLSL, solid view, double sided, no textured solid: 35 fps
GLSL, solid view, no double sided, textured solid: 55 fps
GLSL, solid view, double sided, textured solid: 55 fps
GLSL, solid view, No double sided, no textured solid: peaks at 120 fps

One more subdivision applied to 983k triangles:
GLSL, solid view, No double sided, no textured solid: 78 fps

Another subdivision applied to 3.93 mio triangles:
GLSL, solid view, No double sided, no textured solid: 21 fps

This might explain my bad results with the HD5850, seems very, very dependent on the viewport settings as well. Got to check it right away :wink:

So it might be better to start over, with a .py testing all the modes and benchmarking it, and maybe directly submit the results to an online database.
The question is though, which are the modes most likely to be used and it varies with the task.

and 15 minutes later… :smiley:
HD5850:

GLSL, solid view, double sided, textured solid: 35 fps
GLSL, solid view, no double sided, textured solid: 35 fps
GLSL, solid view, double sided, no textured solid: peaks at 120 fps
GLSL, solid view, No double sided, no textured solid: peaks at 120 fps

One more subdivision applied to 983k triangles:
GLSL, solid view, double sided, no textured solid: 55 fps
GLSL, solid view, No double sided, no textured solid: 55 fps

Another subdivision applied to 3.93 mio triangles:
GLSL, solid view, double sided, no textured solid: 15 fps
GLSL, solid view, No double sided, no textured solid: 15 fps

Conclusion:
My GTX is OC´d so it seems as long as you turn off double sided the Radeon and GeForce are on par, even with textured solid on.
Man this whole thing starts to give me a headache :smiley:

Any python coders want to work on a benchmarking addon?

Graphic card type: Palit gtx560 2gb
Graphics driver version: 270.29
Number of graphic cards in system: 1
CPU type: Phenom II 720 x3
Mainboard type: Gifabyte GA-MA69VM-S2
Operating System: Ubuntu 10.04 x64
System Memory: 8 gb ddr2
Screen Resolution:1280x1024
Notes: 60 fps avg

http://s46.radikal.ru/i113/1105/d3/cb086ab4342at.jpg

Graphic card type: Palit gtx560 2gb
Graphics driver version: 270.29
Number of graphic cards in system: 1
CPU type: Phenom II 720 x3
Mainboard type: Gifabyte GA-MA69VM-S2
Operating System: Ubuntu 10.04 x64
System Memory: 8 gb ddr2
Screen Resolution:1280x1024
Notes: 60 fps avg

http://s46.radikal.ru/i113/1105/d3/cb086ab4342at.jpg

You have my sympathy :), and to add further to your growing pain I might suggest you test one round inside sculpt mode.

I have to test your blend file once I am at my private machine at home with my gtx460, but my previous tests show similar results to yours, with the addition, that in sculpt mode you get an extra boost in framerate. It may be related to my setup, but if not, this could hint at a way to accelerate the viewport experience.

I’ll do your test, I have a gtx 460, but how do I turn on VBO? Edit it’s ok I found it. 105 fps with it enabled 15 without amazing difference. I’ll post screen grab and details shortly

Go to ‘User Preferences/System’ there in the OpenGL Section just below the Mipmaps.

I found out that the sluggishnes during sculpting mostly results from CPU bottlenecking. If you work with complex or tiny brushes the CPU seem so can´t process the vector transformations for the vertices anymore. It´s because it is only singlethreaded.
Multires the default cube 8 times and use the add brush and if you check your CPU load you´ll see that you have a total of 100% usage at a certain multires level no matter how many cores. On a quadcore 400% would be desireable, someone got to make it fully multithreaded like sculptris for instance. To me Blender’s sculpt is a useless tool as it is now. Nah useless sounds too hard… too painful to use productive is maybe better :smiley:

I guess you misunderstood me :slight_smile:

If I start the animation using ctrl-a inside sculpt mode, I get increased frame rates, up to a factor of 2-3. The sluggishness during sculpting is another topic altogether.

seanser what window draw method are you using? cause I have an gtx460 with VBO on and all I get is 30 fps.

Nvidia GTX 460 GLH
Just downloaded latest drivers 27061
1 graphics card
Intel i7 980x
Gigabyte X58A UD3R
Windows 7 64bit pro
12 Gig RAM
screen res 1680 x 1050
fps = 108

http://img35.imageshack.us/img35/976/arexmatst.jpg

Sam, I always turn off smooth view because I hate it.

Stays exactly the same for me.

Tried that resolution just now. stays 75fps with my [email protected]
I know the GLH is OC´d quite heavy, but it can´t be faster. Either your GPU is very different, or the Ci7 accounts for the difference.
You had “GLSL, solid view, double sided, textured solid” ?

I didn’t change anything at all from your scene. I’ve only just downloaded 2.57b today and the only changes I made were to turn off smooth view and enable vbo s. I also turned off in the Nvidia control panel vertical sync just as you suggested. My GPU is is definitely a 460 Goes Like Hell:). I can understand your point though, it shouldn’t be faster than a 470.

I believe Smooth view made the difference there. Smooth view is a bit of a taker.

what do you mean with “smooth view”? where is this option?

EDIT: Found, I changed the value to zero but I don’t see any difference at all, not even in the performance, I don’t know how do you get 105 fps with the same card.

You might have to check in the properties panel (n) if GLSL is on and if [x] textured solid is enabled.
TBH I have no idea if those settings are blender configuration properties or file properties, b/c all my Blenders have it as default.

when I switch go GLSL mode I get about 70% that’s better, but far from 105 fps :-S , maybe is because the resolution, mine is higher.

I gettin an awful feeling that you guys think i’m telling porky pies here’s the full Blender screen. I’m now getting 105 fps, I have a couple of other programs running and since I first ran the test earlier I’ve started configuring 2.57b to my own liking, which basically means I’ve changed the view port colours and enabled the loop tools addon.http://img215.imageshack.us/img215/7659/arexafullscrn.jpg