SPLATONIC: Architectural Support for 3D Gaussian Splatting SLAM via Sparse Processing

HPCA 2026

Webpage | Paper | arXiv

demo.mp4

This repository cantains the official software implementation of SPLATONIC, a sparse and efficient realtime 3D Gaussian Splatting (3DGS) SLAM algorithm-hardware co-design with a novel rendering pipeline for resource-constrained device. Specifically, SPLATONIC proposes an adaptive pixel sampling algorithm with up to 256x computaion reduction as well as comparable accuracy, and address the underutilization in original training process with pixel-based rendering pipeline. SPLATONIC could achieve 14.6x speedup on NVIDIA AGX Orin mobile SoC on average.

This implementation is based on SplaTAM, we will release implementations based on other algorithms in the future.

Installation

To install requirements:

conda create -n splatonic python=3.10
conda activate splatonic
# Example for CUDA 12.8:
pip3 install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu128
pip3 install -r requirements.txt

Note: PyTorch installation varies by system. Please ensure you install the appropriate version for your hardware.

Data Preparation

Download the datasets according to instruction from SplaTAM. The datasets will be stored into ./data by default.

# Replica
bash bash_scripts/download_replica.sh

# TUM-RGBD
bash bash_scripts/download_tum.sh

Usage

Run

Taking Replica as an example, to run SplaTAM with SPLATONIC, please run the following command:

python scripts/splatam_sparse.py configs/replica/splatam.py

To run SplaTAM without SPLATONIC, please run the following command:

python scripts/splatam.py configs/replica/splatam.py

Similar to SplaTAM, you can modify the configs/replica/splatam.py file or use configs/replica/replica.bash. You can replace 'replica' with 'tum' to evaluate TUM-RGBD dataset.

Benchmarking

We evaluate performance by measuring the elapse that pipeline completely execute a forward and backward pass, reported in the script code as Iteration time. We also provide Frame Time that reports elapse the application execute a complete frame tracking/mapping.

Acknowledgement

We sincerely thank 3DGS, SplaTAM, GradSLAM & ConceptFusion and for their open-source code.

Citation

If you find this work helpful, please kindly consider citing our paper:

@inproceedings{huang2026splatonic,
  author = "Huang, Xiaotong and Zhu, He and Ma, Tianrui and Xiong, Yuxiang and Liu, Fangxin and He, Zhezhi and Gan, Yiming and Liu, Zihan and Leng, Jingwen and Feng, Yu and Guo, Minyi",
  title = "SPLATONIC: Architectural Support for 3D Gaussian Splatting SLAM via Sparse Processing",
  year = "2026",
  booktitle = "Proceedings of the IEEE International Symposium on High Performance Computer Architecture",
}