With my little and dangerous google.com knowledge I would go for a 8 GPUs build with 4 x NVIDIA GeForce 9800 GX2 to get estimated 3 500 000 000 hash/sec.
Spec has been developed by "researchers at the University of Antwerp in Belgium" to "perform their tomography work very efficiently" (http://www.dvhardware.net/article27538.html).
"The main reason why they configured an AMD system is because they couldn't find a motherboard for the Intel platform that could fit four GeForce 9800 GX2 graphics cards. Another interesting note is that this system doesn't need SLI, their application uses the NVIDIA CUDA programming model which makes all eight GPUs work in parallel. The researchers say they don't need SLI during a reconstruction as every graphics card communicates directly with the CPU, no inter-GPU communication is needed."
Here's a look at the specifications
- AMD Phenom 9850 processor + Scythe Infinity CPU cooler
- 4x MSI 9800GX2 graphics card
- 4x 2GB Corsair Twinx DDR2 PC6400 memory
- MSI K9A2 Platinum motherboard
- Samsung Spinpoint F1 750GB HDD
- ThermalTake Toughpower 1500W Modular PSU
- Lian-Li PC-P80 Armorsuit case
- Windows XP 64-bit