SESHAT: Systematic Energy & Seed-Provenance Harness for Algorithmic Tradeoffs

SESHAT is a reproducible benchmark suite for modular exponentiation on embedded Arm and RISC-V system-on-chip (SoC) platforms typical of Internet of Things (IoT) and unmanned aerial vehicle (UAV) gateways. SESHAT evaluates binary square-and-multiply, the constant-time Montgomery ladder, and fixed-window methods under pinned CPU governors on Raspberry Pi 5 (Cortex-A76), PYNQ-Z1 (Cortex-A9), and VisionFive 2 (RISC-V JH7110), reporting latency (median, 95% confidence intervals), idle-subtracted energy, and hardware counters. This work targets thesingle-operationregime (B=1, whereBis thebatch size) and introduces a compact model with thresholdB⋆= ⌈tpre/(tbin−tfw)⌉ that predicts when fixed-window amortizes its precomputation. Measurements and the model agree: across 1024–4096-bit moduli, thereis no crossover at B=1; fixed-window becomes preferable only onceB≥B⋆. A brief microarchitectural rationale indicates that algorithmic precomputation costs primarily explain the ordering, with branch prediction and cache locality playing secondary roles across architectures. Seed-provenance logging using Global Positioning System (GPS) time and a true random number generator (TRNG) is audit-oriented and performance-neutral.