The problem is in D3 engine it is not Crysis nor latest Unreal so models should be simple (maybe not as simple like low poly from old 3d games) and use bumpmaps to make them look more advanced. If You make models with too much points on grid and details (that can be simple show as textures) game will slow down. D3 should works fine on latest computers and those bit older ones. For example it works very good on my laptop with Vista, intel graphic card and 1gb ram (medium quality settings, some times when there are many monsters it slow down a bit).
The most demanding aspect of the Doom3 engine is the lighting system. The way it works is that it extrudes geometry from the silhouette (or the whole object. I'm not sure) of objects for each light in the scene. So let's say there are 4 monsters and 3 lights in a room. The calculated shadow volumes will then be of 3 * 4 * monster triangles. With a detailed model that would be a lot of triangles. However, it's possible to have a simpler model for the shadow generating purposes. With that, we can minimize the impact of higher polycount. The new models in Phobos are generally a tad more detailed than the ones you see in Doom3. We're not going all the way with that, though, as there would be a gap in quality between the retail stuff and our stuff.