Recently, a team from the Computer Network Information Center of the Chinese Academy of Sciences developed HIP-DFPT, a large-scale quantum perturbation computing software for heterogeneous GPU clusters. Targeting the irregular workload characteristics in quantum perturbation calculations, the software proposes methods such as micro-task reorganization, GPU memory optimization, and multi-stream asynchronous pipelining. It reorganizes a large number of fine-grained computing tasks into micro-task units more suitable for GPU execution, improving thread utilization and data access efficiency while reducing communication and data transfer overhead.

Quantum perturbation theory is an important theoretical tool for studying the response properties of materials and molecular systems. It can reveal microscopic physical characteristics such as phonon spectra, polarizability, and Raman spectra, and has significant application value in fields such as materials science, condensed matter physics, and molecular simulation.

The research team proposed an AI-based performance modeling method for complex branching code, which can quickly predict task execution time by sampling only a small amount of runtime data. They also designed a hybrid multi-level load balancing algorithm to achieve coordinated optimization of dynamic scheduling between nodes and within GPUs. Experimental results show that HIP-DFPT can scale to 8192 domestic GPUs and support full-electron precision simulations for systems with up to 200,000 atoms, with a 70% improvement in atomic throughput. This validates the software's parallel scalability and engineering application potential in large-scale quantum perturbation calculations, providing a new technical pathway for intelligent optimization of complex scientific computing programs on heterogeneous GPU clusters.

It is reported that the related research results have been accepted by IEEE Transactions on Parallel and Distributed Systems (a CCF Class A journal).