Vertex throughput on modern GPUs

I've read that the optimal number of triangles in order to maintain 60fps that the GPU can process is something between 1500 and 4000.

Found this interesting link:

http://nuclexframework.codeplex.com/wikipage?title=PrimitiveBatch&referringTitle=Nuclex.Graphics

But according to this presentation about nvidia http://origin-developer.nvidia.com/docs/IO/8230/BatchBatchBatch.pdf?q=docs/IO/8230/BatchBatchBatch.pdf, the question is not how many triangles per batch, but how many batches per frame!

Following that last link, the conclusions that can be drawn are:

- Number of vertexes per batch: whatever between 1 and 10.000.
- Number of batches per frame (CPU bound): 25k * GHz * Percentage_available_for_draw_calls/Framerate
- Bad performance at a small number of vertexes per batch could mean:
- GPU vertex engines are idle but the triangles itself could be expensive
- CPU bound --> add triangles for free!
- GPU idle: add complicate pixel shaders for free!