limbo018/DREAMPlace

What it solves

DREAMPlace is a high-performance toolkit for VLSI (Very Large Scale Integration) placement, which is the process of determining the optimal physical locations of components on a semiconductor chip. It addresses the computational bottleneck of traditional CPU-based placement tools by leveraging the analogy between nonlinear VLSI placement and deep learning training, allowing the process to be accelerated via GPUs.

How it works

By treating the placement problem as a deep learning optimization problem, DREAMPlace utilizes deep learning toolkits (like PyTorch) to perform global and detailed placement. It can run on both CPUs (multi-threaded) and GPUs, using GPU acceleration to achieve significant speedups in global placement, legalization, and detailed placement (via the integrated ABCDPlace).

Who it’s for

It is designed for hardware designers, chip architects, and researchers in electronic design automation (EDA) who need to place millions of components on a chip efficiently and with high precision.

Highlights

• Massive GPU Acceleration: Achieves over 30x speedup in global placement and legalization compared to CPU-based tools like RePlAce.
• Integrated Detailed Placement: Includes ABCDPlace, providing up to 16x speedup over sequential placers like NTUPlace3.
• Advanced Optimization: Supports timing-driven placement, momentum-based net weighting, and 2-stage macro placement flow.
• Broad Input Support: Compatible with LEF/DEF and Bookshelf formats.
• Timing Analysis: Integrates OpenTimer and HeteroSTA for GPU-accelerated static timing analysis.