What's the real SP/DP performance of K20m?

Accelerated Computing CUDA CUDA Programming and Performance
1

NVIDIA announces that K20m’s DP/SP performance is 1.17/3.52 Tflops.
As we known, K20m has 13 SMX(2496 CUDA cores). These cores work at 0.705GHz.

I calculate peek performance like this:
DP performance: 2496 * 0.705 = 1.759 Tflops
SP performance: 2496 * 0.705 * 3 = 5.279 TFflops.

These results are far more than 1.17 and 3.52.

Why?
Thanks.

2

SP perf= 13 (SM) * 192 (sp cores per SM) *.705 (clock) *2 (FMA) = 3.52 TFlops
DP perf= 13 (SM) * 64 (dp cores per SM) *.705(clock) *2(FMA)=1.17 Tflops