Reading the Wiki on efficiency in content creation, I see we are only supposed to use a normal map on the first LOD level. How do we do that without using another material? And anyway the normal map is there, so why not just use it at all LODs?
That sounds like pretty old advice, probably dating back to 2009 or earlier. The objective was probably to reserve higher-cost materials for geometry that was close to the observer, and use cheaper materials in the distance where the difference isn't as noticeable. That's possible "without using another material" because you're probably already switching materials at some point in your LOD chain, so ensuring that the new material is not normal-mapped is free. However, it's definitely not advice that I'd recommend today:
* The cost of normal mapping is entirely trivial compared to everything else we're doing in the TANE shaders. There is a slight increase in texture bandwidth requirements, but given that we're talking about distant LODs, you're probably either going to be using smaller texture or a low MIP anyway, so this doesn't amount to much.
* The material savings are per-fragment; not per-polygon, per-vertex or per-mesh. Since there are typically far less fragments on a distant object, you're unlikely to save much by optimizing the material. An exception would be if there was massive overdraw (which should generally be avoided anyway) or if the majority of the screen was occluded by thousands of repeats of the object in question (eg. a tree).
* With modern techniques, the benefits of normal mapping far outweigh the costs. A non-scientific example: if you can save 1% performance but it costs you 50% visual quality, that's not a good tradeoff. Worse, losing the normal map is likely to mean that you'll require more vertices and polygons to compensate visually, and that WILL cost you noticeable performance.
Short version, don't do this any more. If you'd like to point me to the relevant section on our wiki, I'll update it (or you can).
chris
!!
