We are building an in-house 32-core cluster with two Nvidia TESLA GPUs. This will enable us to complete 5-7 rotor revolution simulations in 2D in under 4 hours.
Although the GPUs have much higher memory bandwidth (~3 teraflops @ 32-bit), the CPU (~1 teraflop @ 32-bit) still significantly outperforms GPU in all CFD models (Spalart-Allmaras, k-epsilon, k-omega, transition-SST. We do, however, expect significant performance boost from GPUs in Large-Eddy-Simulation (LES).