Parallel Optimization of NTT for Kyber Key Encapsulation Mechanism Using GPU-Accelerated Plantard Arithmetic

초록

CRYSTALS-Kyber is a key encapsulation mechanism that has been standardized by the National Institute of Standards and Technology. Number theoretic transform (NTT) is used to compute the polynomial multiplication in Kyber, which is also the main performance bottleneck. In this letter, a graphics processing unit (GPU)-optimized implementation of Plantard arithmetic intended for improving the performance of NTT is introduced. Various techniques are proposed to avoid warp divergence and expensive division in Plantard arithmetic, thus achieving a faster implementation. Experimental outcomes reveal performance improvements of 6.57%, 6.88%, and 6.76% for NTT, as well as 2.18%, 2.67%, and 3.24% for inverse NTT (INTT), when compared to Lee et al., for Kyber-512, Kyber-768, and Kyber-1024, respectively. These results underscore the efficient and scalable implementation of Kyber, rendering it particularly suitable for the Internet of Things platforms where high throughput and moderate latency is imperative.

키워드