The Ring–Star Network Design Problem (RSP) is an NP-hard combinatorial optimization problem that aims to determine an optimal hybrid topology combining ring and star connections. Traditional approaches have primarily focused on minimizing the overall cost, often overlooking the impact of communication energy on network efficiency. To address this, we propose Fast and Energy-Efficient Ring–Star Network Design (FERN), an enhanced formulation of RSP. The objective is not only to incorporate an energy aware perspective but also to achieve faster computation and improved overall cost performance compared to existing state-of-the-art methods. It is implemented using a Max-Min Ant Colony Optimization (MMACO) framework integrated with post problem specific local searches, namely relocate and 2-opt move. As part of the energy aware formulation, βring and βstar are introduced.
Chemical kinetics simulations often require implicit integration of stiff governing equations. Repeated evaluation of source-term Jacobians is a major computational bottleneck in such solvers. We construct an approximate Jacobian by omitting terms with negligible influence, achieving significant computational savings with minimal accuracy loss. A dynamic Jacobian strategy is further introduced to adaptively switch between approximate and exact Jacobians, balancing accuracy and performance during integration. The methodology is implemented and evaluated on CPU and GPU architectures across multiple reaction mechanisms, demonstrating substantial speedups while maintaining accuracy.
Matrix multiplication is fundamental to deep learning, scientific computation, and graph analytics. The prevalence of sparse matrices in these fields offers significant opportunities to reduce memory and compute costs. optimizing compression formats and algorithms is crucial for improving execution time, memory footprint, and energy efficiency.This paper introduces MATANALYZE, a framework designed to generate efficient hardware accelerators for sparse × dense and dense × sparse matrix multiplication. MATANALYZE takes into account the hardware constraints, including the number of DSPs and BRAMs and the characteristics of the input matrices (sparsity and dimensions) and generates optimized custom hardware accelerator designs. The hardware designs explored by MATANALYZE are based on either row-wise or column-wise product algorithms. Compared to conventional methods, these approaches offer more favorable trade-offs between performance and resource utilization. This paper demonstrates that, for a given workload, selecting the appropriate combination of compression format and multiplication algorithm can lead to significant performance improvements, achieving in some cases more than 1000× speedup compared to naive baseline implementations.
MOCVD process finds extensive application in growing thin films for devices such as solar cells and High Electron Mobility Transistors (HEMTs). The quality of the thin films directly impacts the performance of the device. Therefore the quality of thin films is usually improved though repeated experimental runs in a reactor. Reducing the number of trial and error cycles can save a significant amount of time and resources. In this paper, we present a multiscale computational model to model the growth of GaAs films at both reactor and atomic scales. A Computational Fluid Dynamics (CFD) framework is employed at the reactor scale and a Kinetic Monte Carlo (KMC) framework at the atomic scale is developed to study the growth. We also attempt to couple both frameworks to present a correlation between the characteristics of the grown film and the reactor-scale process parameters.
Probabilistic computing is a quantum-inspired paradigm which can efficiently solve computationally challenging problems. In this work we present a probabilistic computing hardware accelerator for solving the Number Partitioning Problem (NPP) with 1024 fully connected p-bits implemented in an AMD Virtex UltraScale+ FPGA. We also propose a methodology to systematically derive critical algorithm hyper-parameters for fast and accurate convergence. Our accelerator successfully solves NPP instances with problem sizes up to 1000 and consistently reaches the best known solution while achieving accuracy and performance comparable to state-of-the-art.
