The　common　objective　of　very　large-scale　integration　(VLSI)　placement　problem　is　to　minimize　the　total　wirelength,　which　is　calculated　by　the　total　half-perimeter　wirelength　(HPWL).　Since　the　HPWL　is　not　differentiable,　various　differentiable　wirelength　approximation　functions　have　been　proposed　in　analytical　placement　methods.　In　this　paper,　we　reformulate　the　HPWL　as　an　l1-norm　model　of　the　wirelength　function,　which　is　exact　but　nonsmooth.　Based　on　the　l1-norm　wirelength　model　and　exact　calculation　of　overlapping　areas　between　cells　and　bins,　a　nonsmooth　optimization　model　is　proposed　for　the　VLSI　global　placement　problem,　and　a　subgradient　method　is　proposed　for　solving　the　nonsmooth　optimization　problem.　Moreover,　local　convergence　of　the　subgradient　method　is　proved　under　some　suitable　conditions.　In　addition,　two　enhanced　techniques,　i.e.,　an　adaptive　parameter　to　control　the　step　size　and　a　cautious　strategy　for　increasing　the　penalty　parameter,　are　also　used　in　the　nonsmooth　optimization　method.　In　order　to　make　the　placement　method　scalable,　a　multilevel　framework　is　adopted.　In　the　clustering　stage,　the　best　choice　clustering　algorithm　is　modified　according　to　the　l1　-norm　wirelength　model　to　cluster　the　cells,　and　the　nonsmooth　optimization　method　is　recursively　used　in　the　declustering　stage.　Comparisons　of　experimental　results　on　the　International　Symposium　on　Physical　Design　(ISPD)　2005　and　2006　benchmarks　show　that　the　global　placement　method　is　promising.　©　1982-2012　IEEE.