viewport performance

I need some help regarding viewport performance in blender. It seems like it should be higher for my computer but I can’t figure out why or if I am simply expecting too much.

I have a quite high poly model (340 000 faces) and blender shows that it takes up around 250MB of memory which doesn’t seem much. And even though in solid mode manipulating it is smooth, in edit mode it gets choppy. I have a geforce 570 card with 2.5GB VRAM and 16GB system RAM and somehow I think I should be able to work without any lag with even higher poly count meshes. Am I wrong? I tried playing with other files with face counts reaching 3M and then even the solid mode became laggy while blender was reporting only 650MB of memory usage.

Am I missing something here? Can anyone tell me what can be done to improve the performance by playing with some settings within blender or even in windows (windows 7 64bit) or explain the behavior to me if this is how it has to be? I would be very grateful.

The strangest thing for me is that while in cycles rendered view the performance is much better than solid/edit mode. Why is that.

afaik: Blender is supposed to run on old hardware and therefore uses some old (gl_select) stuff which isn’t to fast… There is a patch which improves viewport performance with high poly count. search for gl_select patch.

thanks for the info. I will look into it.

Nvidia Fermi cards (4xx/5xx) have some opengl shortcomings, and do not perform well in opengl at all. (Search on the forum for this - it has been discussed at length.) You can try turning off double-sided lighting for objects and turn on VBOs, but that’s about it. AMD cards perform way better currently - but then you are missing out on CUDA. One solution is to install an AMD/ATI card for good opengl performance, and one nvidia card for good CUDA performance.

One 570 user mentioned he improved viewport performance by installing the extreme-g drivers: http://www.tweakforce.com/

And as for edit mode being laggy: hundreds of thousands of verts is a lot of data to work with in an interactive manner - in edit mode VBOs are turned off, as far as I know.

the gl_select patch improves the speed at which you can select objects in Blender, and does not affect overall viewport performance. It was mainly developed with ATI/AMD cards in mind, because selection grinds down to a halt when dealing with more complex scenes on those cards. With the patch, selection becomes almost instant, even in complex scenes.
However, I have done some testing on Nvidia cards as well (even a Quadro 2000), and even on those cards the gl_select patch works quite a bit faster at selecting objects.

It sucks that only these cards perform badly. Is there any hope the situation will change?

VBO doesn’t seem to have any effect in edit mode so I guess you are right it gets turned off.

Don’t quite know what you mean by double-sided lighting. Are you taking about the solid openGL lights that can be turned on and off by the means of small light bulb icons in the system preferences tab? If so then disabling 2 out of 3 doesn’t improve the navigation in edit mode either.

I will check out the driver. Thanks for that.

As for the gl_select_patch, I have found the thread discussing it but there is no current win64 build there yet. And I am more concerned about the overall navigation than mearly selecting things.

The builds by Bat3a have the gl_select patch built-in. To turn off double sided lighting, open the object data tab (upside down triangle), and uncheck “double sided”. Could you let me know if this improves your opengl performance?

This situation is (most) probably not going to change: Nvidia has Quadros to sell.

Unchecking the double sided option doesn’t seem to have any effect and the option is only there when working in blender internal.

The memory usage indication is either not working as designed or I do not know how it works. I had a scene with 1.6M quads and the panel showed 1.5GB of memory being used which given that my card has 2.5GB leaves more than enough space to render the scene and yet I got some CUDA error with regard to size. Then I reduced the scene to 1.3M quads and the memory jumped to 2.5GB and an another attempt at rendering the scene failed. When I lowered the poly count further to 1M the memory usage stayed the same (2.5GB) and I finnaly managed to get a render. I guess I do not know what purpose this indication serves (especially the additional number in parentheses that follows it).

In that case: no luck for you, sorry :frowning: I have heard/read similar stories from other 5xx owners in regards to opengl performance in Blender - including that even turning off double sided lighting does not seem to work (anymore?).

Blame Nvida.

Oh and the new driver didn’t affect the performance in any way either.
Thanks for all the info you provided though. At least I know what is wrong.

Either it is still the same issue or something with the RC1 build but I have been getting even more severe viewport performance drops in solid mode.

I have a scene with around 150k faces with subsurf turned off in the viewport so not too many and the navigation is close to impossible in solid mode. The strangest thing in it all that the performance dropped from very smooth almost instantly upon adding a small number of verts to one of the models in the scene. Anyone else experiencing this kind of behavior?

The positive aspect in all this is that somehow the textured and material preview in cycles works fine. Can anyone explain it to me?

http://www.gouptoday.info/avatar3.jpgThe strangest thing for me is that while in cycles rendered view the performance is much better than solid/edit mode.