Ello probably gave unvoluntarily the right explanation:

If you send 1000 models into the rendering pipeline, CPU and GPU can run in parallel, the CPU preparing the next rendering steps while the GPU is processing the pipeline.

However one extremely large object will need the CPU to wait for the GPU, thus wasting processing power. This is the unbalanced pipeline case that should also be avoided by level design.

The latter case is only slightly slower, but due to the 60 Hz lock you'll see a huge difference in fps.

However, do not use this result as a level design rule! In a normal level, few large objects always render faster than many small objects. Both methods tested here are equally bad. Try to divide the polygons among 30 objects.

As to the negative bps, I'll download the model and look into that problem.