ClusterSim

ClusterSim is a cycle-accurate simulator for modeling modern GPUs executing CUDA programs. It is built on top of GPGPU-Sim and extends its functionality to support new hardware and programming features.

---

New features compared to GPGPU-Sim 4.0

• Group cores into GPCs with gpgpu_n_cores_per_gpc parameter
• New Special registers: %cluster_ctaid, %cluster_ctarank, %cluster_nctaid, %cluster_nctarank...
• New Instructions: mapa, barrier.cluster
• New directives: .explicitcluster, .maxclusterrank, .reqnctapercluster
• New API calls: cudaLaunchKernelExC, cudaMallocManaged,
• Simulationmodel for the SM to SM interconnect

---

⚙️ Setup

Dependencies

Install required packages:

sudo apt-get install -y build-essential xutils-dev bison zlib1g-dev flex \
    libglu1-mesa-dev git cmake ninja-build clang clang-format

Make sure the CUDA toolkit is in your PATH:

export PATH=$PATH:/usr/local/cuda/bin

---

🔨 Build

The simulator builds as a shared library. CUDA applications linked against cudart (as a shared library) can be run on the simulator by adjusting the dynamic link loader path.

mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j

---

▶️ Run

By default, ClusterSim uses the configuration file at configs/tested-cfgs/SM90_H100/gpgpusim.config.

You can change this via the GPUSIM_CONFIG environment variable.

# If not already in rpath:
source enable_simulator.sh

# Verify CUDA program links to simulator’s libcudart
ldd build/bin/network

# Run a CUDA program
./build/bin/network

# Change GPU config
export GPUSIM_CONFIG=<PATH_TO_CONFIG>/gpgpusim.config
./build/bin/network

---

Notes

If you use ClusterSim in your research please cite

@article{ClusterSim,
    title={ClusterSim: Modeling Thread Block Clusters in Hopper GPUs},
    url = {https://tore.tuhh.de/handle/11420/57345},
    doi = {10.15480/882.15858},
    author={Lühnen, Tim and Behera, Jyotirman and Tripathy, Devashree and Lal, Sohan},
    year={2025}
}

Also cite the original GPGPU-Sim project.