that's what i get when i run crowd.c(bone animated character models) sample with 60k actors on 9600GT, note that this scene has no LOD setup and no instancing. If instancing is going to only give me 10 more fps it's not really useful. Also no one would need this much animated actor on screen at the same time. For vegetation, as others also stated, you would not want same tree model all over the forest. for grass and clutter sprite instancing is supported. Say if you are going to use 20 of each tree model in the scene and only 3 of same tree will be visible on screen, instancing would not make too mush difference on performance.

And still, all models are exactly the same.
And now show me a comparison of the same very complex model rendered five times with instancing and without. I am sure that the instancing one will only be slightly faster, if not even slower. As I already posted before, I am not sure if I understand the concept behind instancing correctly, but if I do, it basicly means that it reduces the amount of data pushed to the graphics card if the same object is rendered again and again. Fact is, that this part of the rendering, isn´t really a big problem anymore with todays hardware. A bigger problem actually is to feed the graphicscard with enough data, so that it doesn´t have to wait for the CPU to push new data.

Btw, if I am writing bullshit, please correct me, as I am really willing to learn more.

But when I tried the demo, where I linked the video with instancing and 60000 models no framerate problem, but without instancing the framerate broke down!!!

Are you actually reading what other people write or are you going to throw youtube links in for the next few posts?

I am reading other people posts!
I think model-instancing is very important. As an example the game anno1404. Without model-instancing this game would not been possible like it is, the developers said. There are big cities, big forests etc. In this game many objects are the same, special the vegetation but it looks good! First they tried without instancing and with just some hundred models the framerate was to much less. So they added model-instancing. And you all forgot the shadowsystem! with PSSM shadows every model is rendered 4 times or more. If people use model-instancing in the right way, the difference between with and without model-instancing is amazing!

Looks like you really know much about mesh instancing, I'm impressed. Why don't you write JCL a email or visit him so you can show him what he is doing wrong, because he clearly must be stupid like hell if model instancing is such a performance booster but the internal tests showed that it wasn't that effective.
I'm pretty sure that he will give you some money for your efforts

Easy solution to this:

There is a d3d_instancing variable, and there exists some model instancing routine for A8.

So just drop it into the engine, mark the feature as DONE and let the user decide

Exactly, that'd be perfect
As well it'd be another point to hold against other engines.

Instancing is a very common technique especially when many objects are rendered multiple times. I saw this often while playing with different technologies.

You also should keep in mind that instancing is not only to speed up the process. It might be that it will not bring a lot of speed progress with just a few models but keep in mind that instancing is also to have a low memory footprint. The index and vertex buffer exists only once for all these objects. Thus it can also help to render a lot more given the same memory budget.

This was our intention, but it turned out that it was not worth the hassle. The scenes where instancing offered any advantage, in the form of a frame rate increase, were rare in an ABT engine. In many scenes it was even slightly slower, due to the sorting overhead. Thus, almost no one would have used model instancing and we saved the precious implementation time for other features.

We're often testing methods to improve the renderer, but not all get implemented in a release.