Placement and design planning for three-dimensional integrated circuits
Jason Cong,Guojie Luo
2011-01-01
Abstract:3D integration enables an additional dimension of freedom for processor design. It enhances device packaging density and shortens the length of global interconnects; it thus benefits functionality, performance and power of 3D integrated circuits (ICs). A major portion of these advantages comes from the reduction of interconnect complexity in 3D integration, when the interconnect delay plays an important part in 2D performance barriers. In order to exploit these new features that do not exist in 2D circuits, novel techniques must be developed in physical design, including the stages of floorplanning and placement. The die assignment of standard cells and IP macros, signal/thermal through-silicon via (TSV) planning, thermal-awareness and traditional wirelength metric are the challenges that need to be handled in the 3D physical design. In this dissertation, we focus on three aspects of the placement algorithms for 3D physical design: (i) gradient computation of density penalty functions for analytical placement; (ii) analytical 3D placement algorithms and the support of mixed-size designs; (iii) thermal-aware cell and TSV co-placement for 3D designs. As an early attempt, we presented transformation-based 3D placement approaches. These had already demonstrated that a 4-die 3D implementation can reduce the wirelength by as much as 50% compared to a 2D implementation by taking advantage of a significant amount of TSVs to shorten wirelength. Later on, we presented a 3D placement framework, which consists of a 3D floorplanning step and a 3D placement step. The 3D placement step is the focus of this dissertation. It supports a pseudo-3D placement mode (die assignment is fixed) and a real-3D placement mode (die assignment is variable). The 3D floorplanning step is performed on a coarsened netlist; it is also used to determine the die assignment for the pseudo-3D placement mode and is used as an initial solution for the real-3D placement mode. Particularly, (i) we presented an efficient gradient computation of density penalty functions which are applicable to both 2D and 3D density-constrained analytical placement algorithms; (ii) we discussed the formulation of density constraints to capture the area distribution in a 3D placement region with mixed-size support, and presented the bell-shaped area projection approach and the Huber-based local smoothing approach; (iii) we presented a TSV/cell co-placement feature in the pseudo-3D placement mode to relieve the thermal issues at the placement stage. The gradient computation of density penalty functions is a general technique applicable to both 2D and 3D density-constrained analytical placement algorithms. This technique reduces the runtime for gradient computation by a factor of n compared to a naïve approach, where n is the problem size. It enables an accurate and efficient gradient computation, while state-of-the-art analytical algorithms rely on approximations. The experimental results show that it reduces 15% of the wirelength on the benchmarks with large area variations. Based on the efficient gradient computation method, we were able to formulate the 3D area density constraints with a bell-shaped area projection approach and a Huber-based smoothing approach to measure overlaps. These approaches achieve 30% shorter wirelength with slightly fewer TSVs than the transformation approach with the best wirelength; these approaches also achieve 30% fewer TSVs with a 10% shorter wirelength than the transformation approach with fewest TSVs. The mixed-size designs are supported with a multi-stepsize scheme in the analytical solver; this reduces the wirelength by 27% on average compared to 2D mixed-size implementations. The thermal issues in 3D ICs are challenging due to the stacked structure. Fortunately, the TSVs provide heat flow paths with much greater thermal conductivity than the dielectric layers. We derived a metric to evaluate the TSV and power distribution efficiently as an indirect evaluation of temperature. This metric is integrated into the pseudo-3D placement mode, and is able to reduce the temperature by 30°C with only 5% degradation in wirelength. These algorithms are integrated in the 3D placement stage of a 3D physical design flow, named 3D-Craft and developed at UCLA. The application of 3D-Craft on an open-source microprocessor is demonstrated.