Aaagrh Congratulations!
Well, let's see how fast table based approach could be. There are ~7.5 rounds (out of full 10), each consists of 8 rows, each row requires 8 table lookups in case of 8-bit tables. 8 table lookups take 38 instructions. Let's assume that memory access is instant and ignore pre- and postcalculations. 280x has 2048 ALUs at ~1024 MHz. So 2048*1024/7.5/8/38=920 MH/s. In case of 11-bit tables there will be 6 lookups per row, but either with sbox-ing, or with some exploit of (unknown) internal sbox structure, 16-bit tables seems unrealistic. May be with mixed table sizes for different memories and better load balancing we'll see over 700 MH/s on 280x, but that will be quite a task! Add to this matrix decomposition and may be using GDS region and I'm out of ideas, at least for now.
On the other hand armchair calculations for (not ready yet) bitslice approach show something like 560 MH/s. I hope for good surprises but ready for bad ones
So there is some space till the barrier but no so much