I also use T-rex and having this issue as well, tried low default and high intensity and kernels.
Client side is 32mh~ the highest Ive seen and client pool rate says higher than that usually.
But pool side the rate is almost half that after a couple 6 hour tests. Around 1% stales which are paid at 80%.
T-rex is about 29.5MH but poolside is exactly that as well.
This is on a 1070 setup Linux latest drivers and version plus older ones.

I am using SSL on both Trex and TBM.
Could it make that much of a difference on TBM but not Trex?