NVIDIA cuDSS (Preview) is an optimized, first-generation GPU-accelerated Direct Sparse Solver library for solving linear systems with very sparse matrices. Direct Sparse Solvers are an important part of numerical computing for real-time applications like autonomous driving and process simulation, where increasing complexity and high throughput demand a robust direct solver. cuDSS leverages both the CPU’s sequential computing and the GPU’s parallel computing to solve sparse matrices with only a few non-zero elements per row, resulting in significantly higher performance than CPU-only solvers.

Features

• GPU-Accelerated Solver: Capitalizes on the CPU’s sequential computing and the GPU’s parallel computing to solve sparse matrices efficiently.
• Core Functionality Support: Solves sparse linear systems on single-GPU, multi-GPU, and multi-node platforms, including support for refactorization in cases with multiple systems, and different reorderings and types of matrices.
• Optimized for NVIDIA GPUs: Supports all NVIDIA GPUs, Pascal and newer, allowing integration of direct sparse solvers across various NVIDIA-powered platforms. Benefits from the Grace Hopper Superchip architecture.
• High Performance: Achieves significant performance gains compared to CPU-based Direct Sparse Solvers.

Benefits

• Enhanced Performance: Delivers significantly higher performance than CPU-only solvers by leveraging both CPU and GPU computing.
• Scalability: Supports single-GPU, multi-GPU, and multi-node platforms, making it suitable for a wide range of applications.
• Stability: Built to be stable regardless of matrix size, ensuring reliable performance.
• Versatility: Optimized for NVIDIA GPUs, allowing integration across various NVIDIA-powered platforms.