As one typical AI-for-Science application, machine learning interatomic potentials (MLIPs) have revolutionized the representation of potential energy surfaces. MLIPs can be categorized into specialized MLIPs, which prioritize high accuracy for specific systems, and pretrained MLIPs, which emphasize generalizability across chemical spaces. Specialized MLIPs and pretrained MLIPs differ in the dataset to be trained, model capability (parameters), the training workflow, and the workload in molecular dynamics. We review different high-performance computing (HPC) optimization techniques for training and inference that specialized MLIPs and pretrained MLIPs tend to prefer. For example, from the perspective of the training dataset, we investigate the load balance strategies, which are critical for pretrained MLIPs to enhance scalability. From the perspective of model parameters, we indicate that specialized MLIPs can benefit from curvature-aware optimization algorithms given their moderate model size. We remark that advances in HPC are not merely engineering improvements but play a key role in faster iteration of MLIPs, broader applicability, and sustained progress in MLIP development.

The growing demand for semiconductor devices simulation poses a big challenge for large-scale electronic structure calculations. Among various methods, the linearly scaling three-dimensional fragment (LS3DF) method exhibits excellent scalability in large-scale simulations. Based on algorithmic and system-level optimizations, we propose a highly scalable and highly efficient implementation of LS3DF on the Sugon supercomputer, a domestic supercomputer equipped with deep computing units. In terms of algorithmic optimizations, the original all-band conjugate gradient algorithm is refined to achieve faster convergence, and mixed precision computing is adopted to increase overall efficiency. In terms of system-level optimizations, the original two-layer parallel structure is replaced by a coarse-grained parallel method. Optimization strategies such as multi-stream, kernel fusion, and redundant computation removal are proposed to increase further utilization of the computational power provided by the heterogeneous machines. As a result, our optimized LS3DF can scale to a 10-million silicon atoms system, attaining a peak performance of 34.8 PFLOPS (21.2% of the peak). All the improvements can be adapted to the next-generation supercomputers for larger simulations.