ÎçÒ¹AV

The O’Donnell Institute scientists are developing several AI-enabled data compression algorithms and testing them on SMU’s HPC platforms. Data compression is a key research area focused on reducing the size of digital data, such as text, images, video, or audio, to enable faster transmission and more efficient storage. Unlike traditional rule-based techniques that depend on fixed statistical models, modern AI/ML methods use neural network and deep learning architectures including Convolutional Neural Networks (CNNs), Autoencoders, Transformers, and Generative Adversarial Networks (GANs) to learn complex patterns directly from large datasets. These approaches achieve higher compression ratios and improved perceptual quality compared to classical algorithms by capturing both local and global dependencies. AI-enabled data compression has gained momentum in recent years due to its impact on optimizing communication, lowering infrastructure costs, and enabling data-intensive applications in fields such as medical imaging, satellite communications, and multimedia streaming.

We are excited to share our latest GitHub release on LLM-Based Text Compression with Arithmetic Coding on Distributed GPU Systems:

This work explores hybrid compression pipelines that combine transformer-based LLMs (BERT, RoBERTa, T5, and Llama) with Arithmetic Coding for reproducible, scalable text compression experiments on HPC systems.

The repository includes:

• Multi-GPU and multi-node workflows
• Deterministic inference pipelines
• SLURM-based HPC execution scripts
• End-to-end fine-tuning and compression implementations

Built and tested on SMU's NVIDIA DGX A100 SuperPOD infrastructure.

Looking forward to discussions around AI systems, compression, distributed computing, and HPC-enabled generative AI research.