BitMoD: Bit-serial Mixture-of-Datatype LLM Acceleration [Paper]

BitMoD is an algorithm-hardware co-design framework for LLM acceleration using bit-serial hardware with mixture-of-datatypes. It supports diverse precision and data types with a flexible accuracy-efficiency trade-off.

This repository contains the source code for reproducing the experiments of our HPCA'25 paper "BitMoD: Bit-serial Mixture-of-Datatype LLM Acceleration".

News

• [2024/11] 🔥 BitMoD is accepted to HPCA 2025 !
• [2024/10] 🔥 We have extended OmniQuant to support BitMoD datatypes.
• [2024/07] 🔥 We have extended AWQ and SmoothQuant to support BitMoD datatypes.

Getting Started

Every folder in this repo is used for a separate set of experiments in the BitMoD paper. Please go to each folder and follow its README to run different experiments.

Perplexity Results

Model	Precision	Quant Method	WikiText2 PPL	C4 PPL
Llama-2-7B	fp16		5.47	6.97
Llama-2-7B	w4g128	AWQ + BitMoD	5.59	7.09
Llama-2-7B	w4g128	OmniQ + BitMoD	5.56	7.06
Llama-2-7B	w3g128	AWQ + BitMoD	6.07	7.64
Llama-2-7B	w3g128	OmniQ + BitMoD	5.86	7.55
Llama-2-13B	fp16		4.88	6.47
Llama-2-13B	w4g128	AWQ + BitMoD	4.96	6.55
Llama-2-13B	w4g128	OmniQ + BitMoD	4.95	6.55
Llama-2-13B	w3g128	AWQ + BitMoD	5.27	6.88
Llama-2-13B	w3g128	OmniQ + BitMoD	5.17	6.84
Llama-2-70B	fp16		3.32	5.52
Llama-2-70B	w4g128	AWQ + BitMoD	3.40	5.57
Llama-2-70B	w3g128	AWQ + BitMoD	3.70	5.77
Llama-3-8B	fp16		6.14	8.88
Llama-3-8B	w4g128	AWQ + BitMoD	6.50	9.32
Llama-3-8B	w4g128	OmniQ + BitMoD	6.45	9.34
Llama-3-8B	w3g128	AWQ + BitMoD	7.79	11.07
Llama-3-8B	w3g128	OmniQ + BitMoD	7.56	11.06
Llama-3-70B	fp16		2.85	6.73
Llama-3-70B	w4g128	AWQ + BitMoD	3.19	6.94
Llama-3-70B	w3g128	AWQ + BitMoD	4.48	7.76

Code Structure

Repo Root
|---- SmoothQuant-BitMoD   # Running SmoothQuant with basic INT and our proposed BitMoD data types
|---- AWQ-BitMoD           # Running AWQ with basic INT and our proposed BitMoD data types
|---- OmniQuant-BitMoD     # Running OmniQuant with basic INT and our proposed BitMoD data types
|---- bitmod-quant         # Weight-only quantization with different precision and data types (e.g. INT, FP, BitMoD)
|---- bitmod-sim           # BitMoD accelerator simulator

Citation

@article{chen2025hpca,
  title={{BitMoD}: Bit-serial Mixture-of-Datatype LLM Acceleration},
  author={Yuzong Chen and Ahmed F. AbouElhamayed and Xilai Dai and Yang Wang and Marta Andronic and George A. Constantinides and Mohamed S. Abdelfattah},
  journal={IEEE International Symposium on High-Performance Computer Architecture (HPCA)},
  year={2025}
}

---

This work is subject to a patent application filed by Cornell University.