Cigma-Sim

Cigma-Sim is a PyTorch-based simulator for bit-sparsity-aware GEMM accelerators. It is designed for architecture-level evaluation with two complementary perspectives:

• Cycle simulation: estimates execution cost under specific PE/array configurations.
• Value simulation: emulates numerical behavior under hardware-style precision and alignment constraints.

Warning

Performance Note: The core framework is specifically designed and optimized to leverage torch.compile in graph mode. Please be aware that executing in eager mode will incur substantial performance degradation.

What It Covers

• Quantized GEMM simulation (quant x quant)
• Floating-point GEMM simulation (float x float)
• Configurable encoding, mapping, windowing, and lane-sharing behavior

Project Layout

• cigmasim/: core simulation library
  • config/: simulation configuration types
  • core/: low-level kernels (cycle counting, alignment, aggregation)
  • simulator/: high-level simulation entrypoints
  • type/: data format and encoding utilities
• example/: model-specific demos
• container/: container definitions

Note

Settings for Bit-Cigma and Bitlet architectures are defined in example as demos.

Requirements

• Python 3.11+
• PyTorch 2.6+

Citation

If you use this simulator, please cite the Bit-Cigma paper:

@article{bitcigma2025,
  author={Zhu, Zixuan and Zhou, Xiaolong and Wang, Chundong and Tian, Li and Huang, Zunkai and Zhu, Yongxin},
  title={Bit-Sparsity Aware Acceleration With Compact CSD Code on Generic Matrix Multiplication},
  journal={IEEE Transactions on Computers},
  year={2025},
  month={Feb},
  volume={74},
  number={2},
  pages={414--426},
  doi={10.1109/TC.2024.3483632},
}