As a simple benchmark for the capabilites of GPGPU software, I've been testing a hotplate simulation. In a 6400x6400 grid, the future temperature of each cell is an average of the cells around it. I've written it in C (single threaded) and it takes 40 minutes to run on an Intel nehelem. I've written it in Matlab and it takes 7 hours. Finally I wrote it in Matlab using the GPGPU plugin Jacket. It took 20 minutes. That's a 2x speed up over the serial code, with all the abstraction and overhead of Matlab. Next is a test in C with CUDA, but I was still very impressed with the performance of GPGPU in Matlab.
In our HPC lab, we see plugins like Jacket to be a valid substitute for translation of calculations from Matlab to C/fortran. While not the most efficient use of hardware, it is very simple to do and can drastically speed up many applications.