ThermalGuard: A Physics-Guided Cross-Layer Thermal Modeling and Compensation Framework for 3D eFlash IMC SoC

3D in-memory computing (IMC) chips provide high density and bandwidth for AI workloads but suffer from thermal instability that degrades weight accuracy and inference reliability. This work proposes a cross-layer framework, ThermalGuard, and validates it on the hybrid-bonded 3D eFlash analog IMC prototype integrating a 12nm logic die and a 28nm eFlash die with 36MB capacity. The proposed ThermalGuard framework consists of: (1) a physics-guided thermal model achieving a kullback-leibler (KL) divergence of 0.088; (2) a temperature-insensitive differential programming (TIDP) scheme that stabilizes weights from -30°C to 60°C; and (3) a temperature-aware training (TAT) algorithm achieving up to 25.66% and 20.26% absolute accuracy improvement at -50°C and 100°C, respectively. With ThermalGuard enabled, ResNet-18 maintains robust inference accuracy on the 3D eFlash IMC SoC across -50°C to 100°C, achieving up to 45.3% (-50°C) and 34.3% (100°C) absolute accuracy improvement. This work demonstrates thermally resilient analog IMC on a 3D SoC, enabling reliable AI inference under dynamic thermal conditions.