Kinda like doing SSL on GPUs, the need for large batches makes it useless for most workloads.
I think that will improve substantially as more and more applications like routing and SSL offload (and video encoding, etc) for CUDA/GPUs begin to take advantage of massive parallelization which has largely been ignored (even in OS multitasking) in general purpose systems. To me, that's the big deal: that it's coming to (from) the desktop (via gaming cards). All the specialized ASICs that drive things could be subsumed into commodity hardware. If we pour effort into getting massive parallel right, and if it winds up in even the cheapest SoCs, that will really be something to cheer for.