NVIDIA actively worked on calculation accelerator tesla last generation, AMD company could not leave this initiative without reciprocal actions. Yesterday were announced two new fireStream
The elder version fireStream 9370 obtained 4 Gb GDDR5 memory and energy consumption levels not more than 225 W. In floating-point calculations the card ensures double precision and the productivity level 528 gigaflops. In calculations with single accuracy the productivity is equal to 2.64 teraflops. For comparison - Tesla c2070 ensures the calculations speed with double precision at 512 gigaflops level, and calculations with single accuracy are made at 1.03 teraflops. In this case, the energy consumption level Of tesla c2070 reaches 247 W. Problem consists in the optimization of software under ATI stream platform.
The second product fireStream 9350 has only 2 Gb GDDR5 memory, but it can fit in one expansion slot. The energy consumption level does not exceed 150 W, but floating-point calculations are 400 gigaflops for double precision, and 2.0 teraflops for the single accuracy. Both products support OpenCL, DirectX 11 and OpenGL.