An earlier thread mentioned that the main reason Rosetta@home is not switching to GPU use is that their program would need a major rewrite to use more than a few of the GPU processor cores. Their algorithm requires that each processor have over 500 MB of memory. How many copies of that algorithm could you fit in the amount of graphics memory most GPU cards have?

Another reason for not having a 64-bit version - what if they've compiled and tested one already, and found that it did not give enough of a speedup to justify adding yet another application version to their server?