Open3DBench

Official implementation of the paper "Open3DBench: Open-Source Benchmark for 3D-IC Backend Implementation and PPA Evaluation". We aim to build a standardized, replicable, and user-friendly framework for 3D-IC PPA evaluation, specifically providing a benchmark for 3D placement techniques.

0. Overview

We provide 8 synthesized netlists in this repository, which can be partitioned and placed in Place-3D (including the two macro placement strategies mentioned in the paper) to generate DEF results.

The DEF results will be processed in OpenROAD-3D for Place-Opt, CTS, Legalization, Routing, RC Extraction, and HotSpot thermal simulation, leading to comprehensive PPA evaluation results.

The above two steps can be done sequentially or separately. Note that due to the randomness in DREAMPlace across different machines (which slightly affects layout and thus PPA results), if you obtain layout results using Place-3D, 100% reproduction of PPA results is not guaranteed. However, our tests show that the impact is minimal and does not affect the conclusions of the paper.

Alternatively, to reproduce the PPA results, we provide 16 DEFs corresponding to the two 3D placement strategies in Table 3 of the paper for 8 designs. By deploying the corresponding version of OpenROAD as instructed below, you can perfectly reproduce the PPA results in the table.

1. How to build

1.1 Build Place-3D

If you want to perform partitioning and 3D placement, you need to install the standard environment for DREAMPlace. We strongly recommend using the official Docker environment provided by DREAMPlace as it is very convenient and easy to use.

After configuring the environment, you need to first download the dataset, unzip it and place it in the Place-3D/ directory. Next, execute the following commands:

cd Open3DBench/Place-3D
mkdir build
cd build
cmake ..
make
make install
cd ..

At this point, Place-3D is successfully configured.

1.2 Build OpenROAD-3D

If you want to perform 3D PPA evaluation, you need to install the corresponding version of OpenROAD. Two installation options are provided below:

Option 1 (suggested if you have already built the OpenROAD environment):

Follow the official OpenROAD build instructions with the commit hash specified in the paper as fbca14c.

Option 2 (suggested if you have never used it before):

Use the pre-built binary (if you choose ubuntu, 20.04 is the only version applicable, as detailed below).

Caution! The binary for Ubuntu 22.04 contains a segmentation fault for RC extraction, which is the problem of the binary package itself. An alternative is to use the Ubuntu 20.04 version. You can pull a docker image for Ubuntu 20.04 and then build OpenROAD and run our code smoothly.

More installation details can be found here.

1.3 Build HotSpot

To run thermal simulation for both 2D and 3D-ICs, you should build HotSpot. First navigate to Open3DBench/OpenROAD-3D/flow/HotSpot, then you can choose whether or not to build with SuperLU (how to build), which will speed up the simulation process. If you have built SuperLU, please modify the Makefile parameter SuperLUroot

SuperLUroot = /path/to/your/build

Then build HotSpot with option SUPERLU=1,

make SUPERLU=1

Or if you want to build without SuperLU,

make SUPERLU=0

2. How to use

2.1 Run Place-3D

Place-3D includes 3D tier partitioning and 3D placement. The initial netlists are already included in the benchmark downloaded earlier. After following the build process above, navigate to the Place-3D directory: cd Open3DBench/Place-3D, where multiple run scripts are provided.

The script run_2D.sh includes commands for running Hier-RTLMP-2D and DREAMPlace-2D (with pre-computed macro placement results for Hier-RTLMP; only global placement is executed here).

The script run_3D.sh includes commands for running Open3D-DMP, and run_3D_tiling.sh includes commands for running Open3D-Tiling.

For specific commands, refer to run_2D_all.sh and run_3D_all.sh, which can be executed directly to run all designs, or you can copy a single line to execute an individual case.

If you successfully execute the steps above, the corresponding DEF output will be located in the path Open3DBench/Place-3D/install/results/. You can proceed to the next step for PPA evaluation.

If you cannot run Place-3D but want to directly reproduce the 3D PPA evaluation results, you can use the DEF files we provide. These correspond to the results of Open3D-Tiling and Open3D-DMP methods on all 8 cases in Table 3 of the paper. They can be used directly for the next evaluation step. Download the results.zip file from this link and extract it to the path Open3DBench/Place-3D/install/. Note that if you have not built Place-3D, you will need to manually create the install folder in the Place-3D directory.

2.2 Run OpenROAD-3D

OpenROAD-3D includes the complete 3D PPA evaluation flow.

Navigate to the directory cd Open3DBench/OpenROAD-3D/flow, where two scripts, run_2D.sh and run_3D.sh, are provided for running the two 2D methods and two 3D methods, respectively.

You can copy any line from these scripts and run it directly in the current directory to execute the corresponding placement strategy and test design.

Here, 2D refers to DREAMPlace-2D, 2D_mp refers to Hier-RTLMP-2D, 3D refers to Open3D-DMP, and 3D_tiling refers to Open3D-Tiling.

Note that the corresponding placement steps in Section 2.1 must be completed beforehand, as the scripts will look for the layout DEF results in the path Open3DBench/Place-3D/install/results/ as input.

The results will be stored in the directories Open3DBench/OpenROAD-3D/logs and Open3DBench/OpenROAD-3D/results.

3. Citation

@article{shi2025open3dbench,
      title={{Open3DBench}: Open-Source Benchmark for {3D-IC} Backend Implementation and {PPA} Evaluation}, 
      author={Yunqi Shi and Chengrui Gao and Wanqi Ren and Siyuan Xu and Ke Xue and Mingxuan Yuan and Chao Qian and Zhi-Hua Zhou},
      year={2025},
      journal={arXiv preprint}
}