why are GPU buffered objects still slow to render?

[mrdoob]

why are GPU buffered objects still slow to render?

[mrdoob]

The matrices get calculated per object whether they change or not... you can try doing mesh.autoUpdateMatrix = false; and updateMatrix() whenever you need them to update and see if that speeds up things...

There are a few other optimisations that could be done right in the Renderer now you mention...

It does make a 30% difference or so. When I have 900 objects, the fps rises from 7 to 10. When I have 10.000 objects, the fps stays below 1 ;-)

Is it really impossible to not send any information to the GPU when the object has not changed at all?

[mrdoob]

I don't really know...

I will try to figure it out...

[supereggbert]

I've found the bottle neck to be the setUniform, you're pushing at least 10.000 16 float transform matrices to the card every frame and it's quite slow with JS at the moment, making state sorting that much more important, unfortunately not much use in this case :-( You could try creating a new material for each cube so the uniforms only have to be set when they're updated but that would probably be a huge memory drain and it would take ages to compile that many shaders. There are a few other options but they get progressively more complicated to implement.

Hmm, that's a pity.

The thing is that I thought/hoped that it would be possible to not push all those matrices to the GPU, because they didn't change. In other words: why can't the GPU just keep rendering all its buffered objects until they change?

[supereggbert]

It is possible on modern hardware but webgl doesn't support it yet. There was talk a while ago about including in webgl v1.0 (even some early demos using it) but I believe it was decided to leave it out for now. I'd suspect it's going to be among the first extensions so might be here in the next few months, at least in the nightlies, fingers crossed ;-)

Kinda hard typing with all my fingers crossed, but if that is what it takes to get it implemented ;-)
Seriously though, I really hope that it gets implemented! I see why it was postponed though, because it is only an issue when there are thousands of objects to be streamed to the GPU.
Anyways, let's just wait for this feature, or for faster streaming! :-)