There are few GPU-s out there, usable for massive parallel computing. Two main platformss are CUDA by NVIDIA and Stream SDK by ATI. NVIDIA re-branded their GTX 280 as TESLA C1060, specially ment dor computations only. It uses same 240 cores platform at same GPU speed, increasing memory to 4 GB per card.
Here is price/performance analysis, prices from www.tigerdirect.com,
less is better
Conclusion is that GTX 280 is best choice, cause price/performance value is good and performance per card is exeptionally high.