ScholarWorks@성균관대학교

초록

Modern computing applications - ranging from cryptography and Monte Carlo inference to reinforcement learning - demand entropy sources with tunable statistical and temporal properties matched to specific workloads. However, most semiconductor-based entropy generators rely on a single dominant physical mechanism, limiting control over stochastic characteristics and temporal dynamics. Moreover, many approaches employ non-complementary metal-oxide-semiconductor (CMOS) materials, limiting large-scale integration. Here, CMOS-compatible entropy source with electrically tunable temporal correlation by rebalancing defect dynamics in foundry-fabricated fully depleted silicon-on-insulator (FD-SOI) transistor is reported. The device hosts two distinct entropy sources: 1) generation-recombination processes from channel defects and 2) carrier-number fluctuations from gate oxide traps. Unipolar gate-pulse stress provides tunability of the entropy source, shifting the dominant mechanism from channel defect-driven Lorentzian noise (long autocorrelation) to oxide trap-driven 1/f noise (short autocorrelation) without increasing noise magnitude. Leveraging this capability, autocorrelation in situ is tuned: strong for momentum building, low for precise actuation, and negligible for correlation-insensitive tasks, and across benchmarks surpasses pseudo-RNG baselines in efficiency and performance. The results demonstrate that, beyond well-studied oxide traps, previously overlooked channel defects in FD-SOI, can be harnessed as entropy source, reframing CMOS transistors as a scalable platform for hardware-based reinforcement learning and stochastic computing.

키워드