A Unified Instruction-Set-Based Hardware Architecture for ML-KEM and CRYSTALS-Kyber

The public-key cryptography, based on integer factorization and discrete logarithms, is facing more and more severe threats from large-scale quantum computers. NIST selected CRYSTALS-Kyber and subsequently standardized its refined version as ML-KEM. This paper presents a unified instruction-set-based FPGA architecture supporting ML-KEM and CRYSTALS-Kyber across all three parameter sets. A custom 64-bit instruction set supports primitive operations and autonomous key generation, encapsulation, and decapsulation. Shared transform, hash/sampling, and codec accelerators execute both protocols without duplicating complete datapaths or requiring FPGA reconfiguration. Dependency-aware scheduling overlaps independent computations, while streaming transfers and operand retention reduce redundant data movement. Pipelined arithmetic and registered interfaces improve timing while preserving protocol-specific sequencing. The implementation achieves post-route internal timing closure at approximately 210 MHz using 13,172 LUTs, 5603 flip-flops, five DSP blocks, and 14 BRAM tiles. Under matched implementation conditions, the unified design saves 49.7% of LUTs, 49.5% of flip-flops, and 50% of DSPs and BRAMs relative to two separate single-protocol builds. Functional and negative tests cover all six configurations, while timing experiments distinguish sampling variation from command latency. Comparison with state-of-the-art FPGA implementations highlights the architecture’s resource-sharing and dual-protocol capabilities.