Thermal issues are critical in 2.5D/3D IC design, and liquid cooling provides an effective solution for heat dissipation. Widely used compact thermal models (CTMs) convert chips into circuit networks for fast thermal simulations. However, current matrix-solving acceleration methods for CTM-derived circuits are inadequate for high-speed iterative transient thermal analysis of large-scale liquid-cooled 2.5D/3D ICs during design optimization. In contrast, the random walk method can provide fast solutions for local nodes in large-scale circuit networks, but it is not applicable to the circuit networks of the CTMs with liquid cooling. In this paper, we propose LCTMwalk, a novel GPU-accelerated random walk method for transient thermal analysis of liquid-cooled 2.5D/3D ICs. To enable random walks on the liquid-cooled CTM-derived circuit network, we replace the voltage-controlled current source model with the diode model. Additionally, we improve the transient analysis by using a time-backward random walk with time-domain path reuse, accelerating the solution of temperature at local circuit nodes. Experimental results show LCTMwalk can solve million-scale cases in only 500 ms, and achieves a 14-22× speedup compared to the state-of-the-art alternating direction implicit (ADI) method with GPU. Besides, LCTMwalk exhibits good generalizability and can be applied to various 2.5D/3D IC structures with high accuracy (error<1 K compared to 3D-ICE).

For highly integrated, thermally constrained chiplets, the design process requires iterative shape optimization, making rapid thermal simulation across varying geometries critically important. Existing deterministic approaches, such as COMSOL and HotSpot require solving large-scale linear systems, incurring expensive computational costs. Stochastic methods suffer from slow convergence, demanding excessive resources for high-precision results. Current neural network (NN)-based methods necessitate retraining upon geometry modifications, limiting adaptability. Meanwhile, neural networks suffer from the absence of provable error bounds, introducing three fundamental risks in practical deployment. We enable the fast solution of heat equations for varying geometries and propose a novel solver that integrates operator learning with stochastic methods. By employing diffeomorphic mapping, our approach addresses the challenge of operator networks in handling shape variations. Furthermore, the network’s predictions guide the stochastic method for variance reduction, which extremely accelerates the traditional stochastic method, while the stochastic results provide error guarantees and corrections for the neural network’s outputs. Extensive experiments show that we achieve a speedup of 10.69-23.04× over commercial field solver COMSOL and a speedup of 5.20-11.87× over the traditional stochastic methods.

With the rapid increase of the transistor number in VLSI, rapidly rising power density and temperatures make heat dissipation a major challenge in IC design and manufacturing. However, conventional deterministic thermal analysis methods have difficulties in obtaining local temperature solutions efficiently, and the existing stochastic method is inaccurate when dealing with thermal analysis problems involving Robin boundary conditions (BCs). In this article, a highly parallelized path integral random walk (PIRW) solver is innovatively proposed for steady-state thermal analysis with mixed BCs, especially Robin BCs. The rigorous calculation of the local time and the Feynman-Kac functional$\hat {e}_{c}(t)$are adopted to accurately handle Neumann and Robin BCs for the first time. Furthermore, based on the PIRW, we propose an accurate and microsecond-level precalculation PIRW (pPIRW) predictor, which precalculates time-consuming random walks, obtains temperatures by simple vector multiplication, and therefore is suitable for proactive thermal management. The pPIRW essentially calculates a partial inverse of large-scale matrices constructed from the finite difference-based compact thermal models (CTMs). Experimental results show that compared with 3D-ICE, the PIRW solver maintains high accuracy with a negligible error within$0.5~^{\circ }$C, achieves$136\times $–$209\times $speedup and$8.53\times $–$11.1\times $storage space reduction with all three kinds of BCs, and decreases to$1\times $–$1.53\times $speedup for lacking the absorbing Dirichlet boundary. The pPIRW further has speed improvement of 2.5e$4\times $–6.7e$6\times $and memory reduction of$36.6\times $–$42.5\times $over PIRW without loss of accuracy. Integrated within a thermal management strategy, the pPIRW predictor can eliminate all thermal conflicts while maintaining the highest working frequency. Meanwhile, pPIRW achieves$29.2\times $speedup and$634\times $memory reduction over the CTM during the offline precalculation stage.