An Index-free Coding Compression Mechanism for Vision Transformers

Vision Transformers (ViTs) have emerged as powerful models in computer vision but their large parameter counts and substantial memory access costs pose challenges for deployment on resource-constrained platforms. This paper proposes an index-free coding compression mechanism for ViTs, designed to enhance compression efficiency and reduce power consumption significantly. Our mechanism incorporates a heuristic clustering-based coding strategy, a dual-aspect coding evaluation method, and a hardware-friendly index-free decoder design. These innovations provide orthogonal improvements to existing compression methods, further enhancing data compression efficiency. Validated through post-layout simulation and synthesis under the TSMC 28nm process, our method achieves a $2 \times$ compression ratio and enhances energy efficiency by $1.99 \times$ compared to existing methods.