Implementing and Measuring ML-KEM on ARMv9-A with SVE2 and SME
Year: 2026
Type: article
Abstract: As quantum computing continues to advance, traditional public-key cryptosystems face increasing vulnerability, necessitating a global transition toward post-quantum cryptography (PQC). A primary challenge for both cryptographers and system architects is understanding how standardized PQC schemes map to the rapidly evolving execution resources of modern processors. ARMv9-A introduces Scalable Vector Extension 2 (SVE2) and Scalable Matrix Extension (SME), which provide wider vector registers, streaming SVE mode, and ZA matrix storage. These features create new opportunities for polynomial arithmetic, but their benefit for latency-oriented cryptographic kernels depends strongly on the realized instruction throughput and data-movement costs. In this work, we present an implementation and study of ML-KEM, the post-quantum key encapsulation mechanism (KEM) standardized by NIST as FIPS 203, on the ARMv9-A architecture. We redesign the polynomial computation pipeline around the vector and matrix execution model. Specifically, we propose two NTT variants tailored to the architectural features of SVE2 and SME: the vector-based NTT (VecNTT) and the matrix-based NTT (MatNTT), which combine layer fusion with architecture-aware data layouts. On Apple M4 Pro, the same-platform comparison is the main result: VecNTT and MatNTT remain slower than optimized Neon on the NTT-heavy kernels and on the full KEMs, even though SVE2 processes four times as many 16-bit lanes. The SVE2/SME kernels accelerate the Kyber reference implementation, with polynomial-vector operations improving by up to about 7.97 ×. An instruction-level study explains the gap: SVE2 mul and sqrdmulh cost about 4.57 × more per coefficient than Neon, so arithmetic throughput, not lane width, dominates NTT latency. To the best of our knowledge, this is the first real-hardware study of SVE2/SME for Kyber/ML-KEM. On Apple M4 Pro, current streaming SVE2/SME does not outperform optimized Neon for this kernel; the paper reports that measurement and the instruction-level account of the gap.
DOI: 10.1145/3863949
Authors: Hanyu Wei, Wenqian Li, Shiyu Shen, Hao Yang, Yunlei Zhao
Institutions: Fudan University, City University of Hong Kong
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7234186244
