Can't use a thread concurrency beyond 4096 on HD7850.

Hey guys, I'm trying to optimize my HD 7850 for LTC mining and I'm having issues. It mines fine at 4096 but only gets 220kh/s. If I set it any higher, my hashrate drops to 30kh/s, no matter what I do, and I finally found out what causes it.

At 4096, GPU-Z reports that I'm using 411MB of the 1GB the GPU has. Fine.
If I go beyond that (to, let's say, 5120), what happens is that GPU-Z reports that only 11MB of the memory is being used, and then a little over 512MB of "dynamic memory" is in use.
What is dynamic memory?
Dynamic memory is system RAM being used by the GPU as RAM. This is so so so so so much slower than using the GPU's RAM and greatly slows down mining.

So why is this happening? Clearly I have enough VRAM to use a higher thread concurrency, what's stopping it? Anyone have any ideas? I'm willing to try anything. I've already tried "setx GPU_MAX_ALLOC_PERCENT 100".

Running Catalyst 13.3 beta with the latest APP SDK.