Accelerated linear algebra for large scale DFT calculations of materials on CPU/GPU architectures with CRYSTAL.

The article focuses on the development and performance assessment of a new hybrid CPU/GPU parallel implementation of linear algebra operations within the self-consistent field (SCF) driver of the Crystal electronic structure package, used for solid-state density functional theory (DFT) simulations. Two GPU porting strategies based on NVIDIA's cuBLAS and cuSOLVER libraries—one prioritizing portability via OpenACC directives and another emphasizing efficiency through explicit CUDA programming—were implemented to accelerate key linear algebra tasks such as matrix multiplication, diagonalization, inversion, and Cholesky decomposition. Benchmarking on systems including α-quartz, a zeolitic imidazolate framework (ZIF-8), and a large mesoporous metal–organic framework (bio-MOF) demonstrated significant speedups over CPU-only parallel versions, with the GPU-accelerated code enabling calculations on very large systems at reduced computational cost; for example, linear algebra tasks for the bio-MOF system completed on a single GPU in comparable time to those on hundreds of CPU cores. The study also discusses the scaling behavior over multiple GPUs, memory management differences between the two porting approaches, and the intrinsic limitations due to workload imbalance and serial code fractions, providing insights relevant for researchers employing high-performance computing in quantum chemistry and materials science.