Faster Scientific Simulations with New FP64 Algorithms in Nvidia cuBLAS (developer.nvidia.com)

Nvidia has announced significant enhancements to its cuBLAS library with the release of CUDA Toolkit 13.4.2, introducing new algorithms that optimize double-precision matrix multiplications (DGEMM and ZGEMM) to run on faster integer Tensor Cores. The new Ozaki-II and 2M algorithms allow high-precision calculations without sacrificing speed, achieving speeds up to 5.5 times faster than traditional methods on specific GPUs. By breaking down double-precision operations into integer computations, these advancements enable researchers in fields like quantum chemistry and materials science to perform simulations more efficiently, making it easier to handle complex workloads that require both speed and accuracy. The implications of these innovations are profound for the AI/ML community and scientific computing at large. As the need for high-speed processing in simulations grows, the ability of cuBLAS to automatically switch to these optimized algorithms without requiring changes to application code simplifies implementation. Coupled with the flexibility of the Automatic Dynamic Precision (ADP) framework, which dynamically selects calculation methods based on accuracy requirements, these developments ensure high performance and precision in demanding computational tasks. With reported performance gains of up to 175 TFLOPS, this advancement positions Nvidia's GPUs as a critical tool for researchers tackling complex problems in molecular simulations and material property studies.