The executing analytical global placers on a flat cell design might give the best placement results, but nevertheless can incur extremely long runtime, especially when the scale of netlists is increasing dramatically nowadays. And in the flatten mode experiments, the results reveal some unexpected situations like several cells with same logical connections are inseparable. Clustering offers an aviliable solution to above questions and give an attractive choice to reduce the scale and complexity of the design, at the same time improve the placement quality. In this paper, we present a new cluster technique called DyadicCluster to consider the internal and external connections between two cells and add the area constraints, so as to combine the most closely two cells, which is more quick and accurate than previous methods. DyadicCluster has been embedded into the global placement process of a nonlinear placer DCNP for large-scale designs. The DCNP runtime explicitly decreases compared to the flatten mode by 44% and the quality improves by 25%. And the half-perimeter wirelength of our placer after detail placement outperforms current state-of-the-art placers Capo, FastPlace, Fengshui and mPL5-fast by 7%, 9%, 1%, and 5% respectively.