Ternip

RTL accelerator by for Ridger Chu's MatmulFree LLM

---

Overview

Ternip is a parameterizable, open-source RTL implementation of a hardware accelerator targeting the MatmulFree LLM algoirhtnm. MatmulFree LLMs replace traditional matrix multiplication with ternary weight operations, enabling significant reductions in compute and memory bandwidth — making them well-suited for hardware acceleration.

This project is licensed under the BSD 3-Clause License and is free to use, modify, and distribute.

Notable Files

File	Description
rtl/ternip/ternip_core.sv	Top-level compute core
rtl/fus/ternip_tmatmul.sv	Ternary matrix multiplication unit
rtl/fus/ternip_rms.sv	RMS normalization unit
rtl/math/	Math modules (sqrt, sigmoid, SiLU, and more)

Tests and build flow are not currently provided but will be made available shortly.

Dependencies

Ternip depends on the BaseJump STL hardware library.

FuseSoC

Ternip is available on the FuseSoC Package Directory. To add it as a library using the GitHub repo directly:

fusesoc library add ternip https://github.com/sifferman/ternip --sync-type=git

Then declare it as a dependency in your .core file:

filesets:
  rtl:
    depend:
      - sifferman::ternip

Citation

If you use this work, please cite the original MatmulFree LLM paper:

@article{zhu2024scalable,
  title   = {Scalable MatMul-free Language Modeling},
  author  = {Zhu, Rui-Jie and Zhang, Yu and Sifferman, Ethan and Sheaves, Tyler and Wang, Yiqiao and Richmond, Dustin and Zhou, Peng and Eshraghian, Jason K},
  journal = {arXiv preprint arXiv:2406.02528},
  year    = {2024}
}

Contributors

See docs/CONTRIBUTORS.