Leveraging SIMD for Accelerating Large-number Arithmetic

Researchers Subhrajit Das, Abhishek Bichhawat, and Yuvraj Patel from IIT Jodhpur and IIT Delhi have introduced DigitsOnTurbo (DoT), a novel approach to accelerating large-number arithmetic by restructuring computations around independent, data-parallel operations. Unlike traditional algorithms that suffer from inherent dependencies limiting SIMD (single instruction, multiple data) adoption on modern CPUs, DoT reimagines the arithmetic pipeline to fully leverage parallel processing capabilities. The results are striking: DoT achieves up to 1.85x speedups for addition and subtraction and 2.3x for multiplication compared to prior SIMD implementations.

When integrated into state-of-the-art libraries, DoT delivers even more impressive gains: up to 4x speedup for addition and subtraction and up to 2x for multiplication. These improvements cascade into real-world performance boosts, including up to 19.3% end-to-end throughput gains for scientific computations and up to 7.9% latency and 5.9% throughput improvements for cryptographic implementations. The work, published on arXiv, targets a critical bottleneck in high-performance computing and cryptography, where large-number arithmetic is foundational.