I can confirm that. I get a 4-16x speedup on a single tesla card. Unfortunately parallelization across multiple gpus is very bad. For that, namd is still the best option.
Dina started to do something cuda-related for saxs.
A bit of the work that was being done to tune AMBER for GPUs in collaboration with NVIDIA http://www.wmd-lab.org/, I head the performance is phenomenal even on a regular desktop machine. Perhaps we can draw some ideas and apply in IMP? There are now also designated GPU servers available from NVIDIA.