Full configuration interaction (FCI) serves as a benchmark tool for evaluating the accuracy of approximate quantum chemical methods. However, its exponential computational cost limits its practical use. The inherently parallel and compute-intensive nature of matrix and vector operations in FCI is well-suited for acceleration on modern GPU architectures. Despite this potential, prior research has only focused on accelerating the matrix and vector multiplication, thus failing to fully exploit the power of GPUs. In this paper, we present a fully GPU-accelerated FCI implementation that uses a CUDA-enabled GPU for the entire computational workflow, achieving end-to-end acceleration. Our implementation incorporates and refines advanced optimization strategies from previous studies, while crucially eliminating redundant computational steps and minimizing CPU-GPU data transfers. To facilitate broad usability and integration, we offer both Python- and C-based implementations. Our C implementation achieves up to a 35.4× speedup over a multithreaded CPU implementation and up to a 15.9× speedup over the current leading GPU-based implementation, respectively. Our Python implementation also delivers substantial acceleration, reaching up to 18.9× and 8.7× speedups over CPU and prior GPU implementations, respectively.
Full configuration interaction (FCI) serves as a benchmark tool for evaluating the accuracy of approximate quantum chemical methods. However, its exponential computational cost limits its practical use. The inherently parallel and compute-intensive nature of matrix and vector operations in FCI is well-suited for acceleration on modern GPU architectures. Despite this potential, prior research has only focused on accelerating the matrix and vector multiplication, thus failing to fully exploit the power of GPUs. In this paper, we present a fully GPU-accelerated FCI implementation that uses a CUDA-enabled GPU for the entire computational workflow, achieving end-to-end acceleration. Our implementation incorporates and refines advanced optimization strategies from previous studies, while crucially eliminating redundant computational steps and minimizing CPU-GPU data transfers. To facilitate broad usability and integration, we offer both Python- and C-based implementations. Our C implementation achieves up to a 35.4× speedup over a multithreaded CPU implementation and up to a 15.9× speedup over the current leading GPU-based implementation, respectively. Our Python implementation also delivers substantial acceleration, reaching up to 18.9× and 8.7× speedups over CPU and prior GPU implementations, respectively.
High Resolution Image Download MS PowerPoint Slide Full configuration interaction (FCI) can provide an exact molecular ground-state energy within a given basis set and serve as a benchmark for approximate methods in quantum chemical calculations, including the emerging variational quantum eigensolver. However, its exponential computational and memory requirements easily exceed the capability of a single server and limit its applicability to large molecules. In this paper, we present a distributed FCI implementation employing a hybrid parallelization scheme with multithreading and multiprocessing to expand FCI’s applicability. We optimize this scheme to minimize the bottlenecks arising from interprocess communications and interthread data management. Our implementation achieves higher scalability than the naive combination of prior works and successfully calculates the exact energy of C 3 H 8 /STO-3G with 1.31 trillion determinants, which is the largest FCI calculation to the best of our knowledge. Furthermore, we provide a comprehensive list of FCI results with 136 combinations of molecules and basis sets for future evaluation and development of approximate methods.
AI-driven methods have demonstrated considerable success in tackling the central challenge of accurately solving the Schrödinger equation for complex many-body systems. Among neural network quantum state (NNQS) approaches, the NNQS-SCI (Selected Configuration Interaction) method stands out as a state-of-the-art technique, recognized for its high accuracy and scalability. However, its application to larger systems is severely constrained by a hybrid CPU-GPU architecture. Specifically, centralized CPU-based global de-duplication creates a severe scalability barrier due to communication bottlenecks, while host-resident coupled-configuration generation induces prohibitive computational overheads. We introduce QiankunNet-cuSCI, a fully GPU-accelerated SCI framework designed to overcome these bottlenecks. It first integrates a distributed, load-balanced global de-duplication algorithm to minimize redundancy and communication overhead at scale. To address compute limitations, it employs specialized, fine-grained CUDA kernels for exact coupled configuration generation. Finally, to break the single-GPU memory barrier exposed by this full acceleration, it incorporates a GPU memory-centric runtime featuring GPU-side pooling, streaming mini-batches, and overlapped offloading. This design enables much larger configuration spaces and shifts the bottleneck from host-side limitations back to on-device inference. Our evaluation demonstrates that our work fundamentally expands the scale of solvable problems. On an NVIDIA A100 cluster with 64 GPUs, our work achieves up to 2.32 × end-to-end speedup over the highly-optimized NNQS-SCI baseline while preserving the same chemical accuracy. Furthermore, it demonstrates excellent distributed performance, maintaining over 90% parallel efficiency in strong scaling tests.
AI-driven methods have demonstrated considerable success in tackling the central challenge of accurately solving the Schrödinger equation for complex many-body systems. Among neural network quantum state (NNQS) approaches, the NNQS-SCI (Selected Configuration Interaction) method stands out as a state-of-the-art technique, recognized for its high accuracy and scalability. However, its application to larger systems is severely constrained by a hybrid CPU-GPU architecture. Specifically, centralized CPU-based global de-duplication creates a severe scalability barrier due to communication bottlenecks, while host-resident coupled-configuration generation induces prohibitive computational overheads. We introduce QiankunNet-cuSCI, a fully GPU-accelerated SCI framework designed to overcome these bottlenecks. It first integrates a distributed, load-balanced global de-duplication algorithm to minimize redundancy and communication overhead at scale. To address compute limitations, it employs specialized, fine-grained CUDA kernels for exact coupled configuration generation. Finally, to break the single-GPU memory barrier exposed by this full acceleration, it incorporates a GPU memory-centric runtime featuring GPU-side pooling, streaming mini-batches, and overlapped offloading. This design enables much larger configuration spaces and shifts the bottleneck from host-side limitations back to on-device inference. Our evaluation demonstrates that our work fundamentally expands the scale of solvable problems. On an NVIDIA A100 cluster with 64 GPUs, our work achieves up to 2.32X end-to-end speedup over the highly-optimized NNQS-SCI baseline while preserving the same chemical accuracy. Furthermore, it demonstrates excellent distributed performance, maintaining over 90% parallel efficiency in strong scaling tests.
We propose an efficient deterministic method to calculate the full configuration interaction (FCI) ground state energy. This method leverages the sparseness of the Lanczos basis vectors that span the Krylov subspace associated with the Hamiltonian to rapidly calculate the lowest eigenvalue of the effective Hamiltonian formed in this basis. By iteratively performing the spanning and diagonalization steps, this method is capable of rapidly reaching chemical accuracy for a variety of strongly correlated molecules, including stretched N2 and C2, in merely several tens of iterations. To accelerate our algorithm, we implement parallelized sparse matrix-sparse vector multiplication, which features a novel hashing method that establishes a one-to-one correspondence between the set of full configuration Slater determinants and a set of consecutive integers. A near-linear acceleration was observed.
GPU-accelerated quantum chemistry programs can dramatically reduce the time required for electronic structure calculations, yet most existing implementations either retrofit GPU kernels onto legacy CPU codebases or optimize individual kernels without addressing workflow-level integration overhead. We present GANSU (GPU Accelerated Numerical Simulation Utility), an open-source quantum chemistry framework written entirely in CUDA/C++ that integrates GPU-accelerated kernels for electron repulsion integrals, Fock matrix construction, and post-Hartree–Fock (post-HF) methods into a unified, GPU-resident execution pipeline. The key design principle is to eliminate host–device data transfers between computational stages by keeping all intermediate data, including density matrices, integral buffers, and Fock matrix replicas, on the GPU throughout the self-consistent field (SCF) iteration, combined with runtime-selectable integral strategies (stored ERI, resolution-of-the-identity, and Direct-SCF) that adapt to system size and available memory. On an NVIDIA H200 GPU, GANSU achieves end-to-end speedups of up to 52× over PySCF for SCF, 45× for MP2 on molecules with up to 470 basis functions, and 44× for FCI, while outperforming GPU4PySCF by up to 34× for FCI, across a range of molecular systems with up to 650 basis functions. The framework further provides analytical energy gradients and geometry optimization with nine algorithms, all operating within the same GPU-resident data flow. These results demonstrate that workflow-aware kernel integration, not just kernel-level optimization, is essential for realizing the full potential of GPU acceleration in scientific computing. GANSU is publicly available under the BSD-3-Clause license.
High Resolution Image Download MS PowerPoint Slide We introduce the first version of GPU4PySCF, a module that provides GPU acceleration of methods in PySCF . As a core functionality, this provides a GPU implementation of two-electron repulsion integrals (ERIs) for contracted basis sets comprising up to g functions using the Rys quadrature. As an illustration of how this can accelerate a quantum chemistry workflow, we describe how to use the ERIs efficiently in the integral-direct Hartree–Fock build and nuclear gradient construction. Benchmark calculations show a significant speedup of 2 orders of magnitude with respect to the multithreaded CPU Hartree–Fock code of PySCF and the performance comparable to other open-source GPU-accelerated quantum chemical packages, including GAMESS and QUICK, on a single NVIDIA A100 GPU.