There is quite a bit of call overhead on mobile platforms as indicated in the various HW related presentations available here:
Your teapot scene has 81 strips and you draw 8 teapots thats a total of 648 draw calls per frame which is a lot of overhead (need to add all the other state change calls on top of this). Generally a game would do about a draw call per object, so realistically for your scene you should aim for a maximum of 8 draw calls - one per teapot. Easiest way to do this is to use indexed triangle lists. If you have good stripped data (very long strips) you could also use a single huge indexed strip but that would probabaly mean adding a couple of degenerate triangles, basically you stitch your 81 strips together into one mega strip.
Doing this will probably result in increased performance.
Btw the same comment applies to your MD2 code as used in the Benchmark. I wrote some MD2 code a while ago and IIRC it uses lots of small strips and fans by default. Again it would probably be better to avoid call overhead by turning this into indexed triangle data and thus a single draw call per MD2 model. I think the MD2 fan and strip sizes are too small to consider stitching them together and you can not do that with fans anyway.
Hope this helps,
K- PowerVR DevRel
