Hi,

I am building an information visualization system similar to a 3D treemap (i.e.: using stacked boxes to represent the proportional size of nodes in a tree). So basically I need to render a lot of 3D boxes. What techniques can I use to improve performance, beside occlusion culling?

Right now, being a prototype, I'm using a very brute force approach: every object has its own VB/IB etc. So when I have to render several hundreds of them, things begin to slow down. I was thinking about having only one VB/IB and just modify the translation and scaling matrices (to obtain different sizes). Is there any other better way?

What techniques can I use to improve performance, beside occlusion culling?

Instancing is the way to go. There are several techniques available, here's an article to get a basic understanding of instancing and different techniques to archieve it (the article is ~7 years old, using modern API like DX11 will ease instancing a lot).

I was thinking about having only one VB/IB and just modify the translation and scaling matrices (to obtain different sizes).

That is definitely the first step you should take. Rendering all from one vertex/index buffer is going to be faster than switching buffers for each object.

After that, you should be in good shape to tackle instancing, which should be ideally suited to your use-case.