A GPU-Accelerated High-Performance Design for CRYSTALS-Dilithium Digital Signature

Our approach leverages the massive parallelism and computational power of a modern Graphics Processing Unit (GPU) to accelerate CRYSTALS-Dilithium's key operations. While post-quantum cryptographic algorithms like Dilithium offer robust security, they face computational challenges in high-throughput scenarios. Dilithium's performance bottlenecks, particularly in polynomial arithmetic and random sampling, have limited its practical deployment. To address these challenges, we leverage GPU acceleration as a promising solution to enhance efficiency. Our approach develops novel GPU-centric algorithms and data structures tailored to Dilithium's requirements. We focus on optimizing Number Theoretic Transforms (NTT), polynomial arithmetic, and random sampling, with an emphasis on batch processing and efficient memory management to maximize throughput. Experimental results demonstrate significant performance improvements over existing implementations. Our GPU-accelerated version consistently outperforms reference implementations, achieving speed-ups ranging from 53.7% to 59.9% for signature generation and 20.0% to 63.8% for verification across various security levels. These results highlight the potential of our approach in enabling efficient post-quantum cryptography for high-performance applications.