cuNNQS-SCI: A Fully GPU-Accelerated Framework for High-Performance Configuration Interaction Selection withNeural Network QQantum States

Researchers introduced cuNNQS-SCI, a fully GPU-accelerated framework that solves a critical scalability bottleneck in neural network quantum state methods for solving complex quantum systems. The system achieves 2.32X speedup over previous CPU-GPU hybrid approaches while maintaining chemical accuracy, demonstrating 90%+ parallel efficiency across 64 GPUs.

cuNNQS-SCI represents a meaningful advancement in computational quantum chemistry by eliminating architectural constraints that previously limited problem scale. The hybrid CPU-GPU design of existing NNQS-SCI implementations created fundamental bottlenecks: centralized CPU-based deduplication caused communication overhead, while host-resident configuration generation imposed prohibitive computational delays. By shifting these operations entirely to GPU execution with distributed deduplication algorithms and specialized CUDA kernels, the new framework removes these constraints and enables researchers to tackle larger quantum systems that were computationally infeasible previously.

This work addresses a long-standing challenge in scientific computing where algorithmic improvements hit practical walls due to architectural limitations. The integration of GPU-side pooling, streaming mini-batches, and overlapped offloading demonstrates sophisticated systems design that manages GPU memory constraints while maximizing throughput. The achieved 2.32X speedup on A100 clusters while preserving accuracy validates that the optimization approach maintains the method's reliability rather than sacrificing fidelity for speed.

For the broader AI and scientific computing ecosystem, this illustrates the continued importance of specialized hardware acceleration for domain-specific problems. Organizations developing quantum simulation tools, materials science research teams, and pharmaceutical companies relying on quantum-based drug discovery stand to benefit from reduced computational timelines. The strong scaling performance suggests these improvements scale across larger GPU clusters, making previously intractable simulations accessible within reasonable timeframes. Researchers using NNQS methods gain immediate practical benefits, while the architectural patterns employed could inform similar optimization efforts in other GPU-accelerated scientific computing domains requiring global coordination and memory-intensive operations.

• →cuNNQS-SCI eliminates CPU-GPU hybrid bottlenecks through fully GPU-accelerated architecture with distributed deduplication and specialized CUDA kernels.
• →Framework achieves 2.32X speedup over optimized baselines while maintaining chemical accuracy on NVIDIA A100 clusters.
• →Strong scaling demonstrates over 90% parallel efficiency across 64 GPUs, indicating effective distributed performance.
• →GPU memory-centric runtime design with streaming and overlapped offloading enables larger configuration spaces than single-GPU memory allows.
• →Advancement enables larger quantum systems to be solved computationally, accelerating materials science and drug discovery research timelines.