← 研踪 PaperTrailarXiv:2611.23690 · v1
A Probabilistic Behavioral RTL Simulation Framework for Probabilistic Compute Fabrics
首次提交 2026-11-18 09:33
版本更新 2026-11-18 09:33
雷达观测 2026-11-19 19:08
摘要
Coupled oscillator Ising hardware can produce different solution qualities across repeated runs of the same problem, while a deterministic behavioral model used for the design yields a single outcome under fixed conditions. Capturing this stochastic variability during design time is essential for evaluating how closely simulation reflects silicon behavior. This work presents a probabilistic behavioral RTL simulation framework that introduces Gaussian timing variation into coupling delays. This variation changes oscillator timing during evolution, producing different outcomes across runs while retaining the existing chip control and readout blocks. The framework is evaluated against measurements from two chips: a 28-nm CMOS chip with 66 spins and a 16-nm FinFET chip with 49 spins, using 250 problems run 100 times repetitively. Compared with deterministic simulation, the model reduces the average absolute gap between sorted simulation and silicon outcomes by 14.0% and 28.4% on the two chips, respectively. These results demonstrate that a simple behavioral timing variation model can be a useful tool for designing and verifying probabilistic computing chips.
作者备注:5 pages, 5 figures
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-18 09:33 · 最新观测版本
A Probabilistic Behavioral RTL Simulation Framework for Probabilistic Compute Fabrics
Coupled oscillator Ising hardware can produce different solution qualities across repeated runs of the same problem, while a deterministic behavioral model used for the design yields a single outcome under fixed conditions. Capturing this stochastic variability during design time is essential for evaluating how closely simulation reflects silicon behavior. This work presents a probabilistic behavioral RTL simulation framework that introduces Gaussian timing variation into coupling delays. This variation changes oscillator timing during evolution, producing different outcomes across runs while retaining the existing chip control and readout blocks. The framework is evaluated against measurements from two chips: a 28-nm CMOS chip with 66 spins and a 16-nm FinFET chip with 49 spins, using 250 problems run 100 times repetitively. Compared with deterministic simulation, the model reduces the average absolute gap between sorted simulation and silicon outcomes by 14.0% and 28.4% on the two chips, respectively. These results demonstrate that a simple behavioral timing variation model can be a useful tool for designing and verifying probabilistic computing chips.
查看此版本解读 →查看此版本原文 ↗