CNN-inspired Analytical Global Placement for Large-scale Heterogeneous FPGAs
TimeWednesday, July 13th1:53pm - 2:15pm PDT
Location3007, Level 3
Event Type
Research Manuscript
Physical Design and Verification, Lithography and DFM
DescriptionThe fast-growing capacity and heterogeneity are challenging for FPGA global placement. This paper presents a CNN-inspired analytical placement algorithm for heterogeneous FPGAs. We compute a density penalty function by a forward propagation with fully connected layers and a gradient by a backward one with a specialized convolution operator. With FPGA heterogeneity, vectorization plays a vital role in self-adjusting the density penalty factor and learning rate. Further, a region-aware virtual wirelength model contributes to site constraints by establishing connections between cells and their nearest available legal regions. Finally, we adopt weight-aware gradient preconditioning and formulate an effective objective function.