cuNNQS-SCI: GPU-Accelerated Framework for Neural Network Quantum State Configuration Interaction
The introduction of cuNNQS-SCI marks a significant advancement in tackling computational challenges associated with neural network quantum state methods. This innovative, fully GPU-accelerated framework focuses on the Selected Configuration Interaction technique, renowned for its precision in addressing the Schrödinger equation within intricate many-body systems. While AI-driven approaches have demonstrated remarkable potential, the earlier NNQS-SCI method struggled with scalability due to its hybrid CPU-GPU setup. Issues arose from centralized CPU-based global de-duplication, which caused communication delays, and host-resident coupled-configuration generation that incurred excessive computational costs. By implementing a distributed, load-balanced global de-duplication algorithm, cuNNQS-SCI aims to reduce redundancy and communication burdens, facilitating its use for larger systems. This work is detailed in the preprint arXiv:2604.15768v1.
Key facts
- cuNNQS-SCI is a fully GPU-accelerated framework for Selected Configuration Interaction.
- It addresses bottlenecks in the NNQS-SCI method for neural network quantum states.
- The previous hybrid CPU-GPU architecture had scalability constraints.
- Centralized CPU-based global de-duplication caused communication bottlenecks.
- Host-resident coupled-configuration generation induced prohibitive computational overheads.
- The new framework integrates a distributed, load-balanced global de-duplication algorithm.
- It aims to minimize redundancy and communication overhead.
- The preprint is identified as arXiv:2604.15768v1.
Entities
—