That's not always true though. There is a significant difference using transparency modes 3 and 6 with transparency and instancing. There is quite a difference with ghosting in almost every mode. This is very interesting to me, as I would have expected similar performance to the alpha tests.
Here's the troubling thing for me - if you do the same tests without alpha mapping then there is very little difference in performance between texture sizes. With alpha mapping the texture sizes make tons of difference. Shouldn't it use mipmapped versions of the alpha mapped textures and automatically balance the speed? It seems to me that DBP does not use mipmaps with alpha transparency. That's certainly not good.
A fantastic benefit of this test is that I can further optimize Geisha House. I use 512x512 textures for a lot of plants. Like I said earlier, without alpha mapping it wouldn't make any difference, the only concern being video memory. Since it is a concern I'll reduce my texture sizes.