Challenges of CAD Development for Datapath Design

In many high-performance VLSI designs, including all recent Intel microprocessors, datapath is implemented in a bit-sliced structure to simultaneously manipulate multiple bits of data. The circuit and layout of such structures are largely kept the same for each bitslice to achieve maximal performance, higher designer productivity, and better layout density. There are very few tools available to automate the design of a general datapath structure, most of which is done manually. Datapath design (from RTL to layout) very often takes a significant amount of human resources in a project. The design is becoming more complex and demanding as the clock frequency is reaching 1GHz, and the process technology is getting to 0.15um and below. Issues with signal integrity, as well as leakage current, are much more significant now as VCC and VT continue to be reduced and current density increases. Elaborate analyses on noise and power are needed for future designs, beyond the already complex timing, reliability, and functional correctness analysis tasks. The burden on CAD tools to support the high-performance microprocessor design is bigger than ever. This paper reviews the general approaches used in the industry to design datapaths from RTL to layout with the difficulties and issues encountered. We propose a new design workflow and a set of tools to improve overall designer productivity, while meeting all other constraints. A description of these tools to support the next generation of microprocessor design is also presented. Our proposed flow allows a designer to choose a design methodology ranging from a fully automated one to a custom one, to a flexible mix of the two. We present a new paradigm of early binding that considers the impact of circuit and layout during RTL design. We also strive to preserve RTL regularity during the circuit and layout design to improve time-to-market. Finally, we present some results on actual design blocks with the proposed tools and workflow, and we suggest future areas for further research.