NVIDIA’s CUDSS is a new solver technology to improve engineering and science computing.

NVIDIA has announced the latest development of CUDSS, the Sparse Direct Solver Library, aimed at improving engineering and science computing. The new versions of CUDSS V0.4.0 and V0.5.0 provide practical performance improvements and useful features to provide tools for data centers and other computing environments.

Main functions of CUDS V0.4.0 and V0.5.0

CUDSS V0.4.0 introduces performance improvements to solve the acquisition and steps with new features such as memory forecast API, automatic hybrid memory selection and variable batch support. Version 0.5.0 adds a favorable host execution mode to smaller matrices and optimizes performance through hybrid memory mode and host multi -threading to further improve these features.

Improving performance and usefulness

Memory prediction API is important for users who need to expect devices and host memory requirements before entering a memory -intensive stage. This helps the scenario where the device memory can be insufficient, so the user can activate the hybrid memory mode with a better efficiency.

CUDSS v0.4.0 also supports non -uniform batching processing to improve performance by accepting various matrix dimensions and rare patterns. In V0.5.0, host multi -threading is introduced, allowing you to run tasks like rearrangement in multiple CPU threads more efficiently.

Significant performance improvement

Updates in CUDSS V0.4.0 and V0.5.0 provide notable performance improvements on various workloads. Version 0.4.0 uses a high -density BLAS kernel when the triangle is dense, accelerates the acquisition and solves the steps to speed up by the permutation of the matrix structure and finance.

In addition, V0.5.0 optimizes hybrid memory mode so that internal arrangements can be resident in the host, which is particularly effective in NVIDIA Grace -based systems due to the high memory bandwidth between the CPU and GPU.

Hybrid Run Mode

The use of the hybrid execution mode introduced in V0.5.0 allows you to run a part of the calculation in the host, which reduces the overhead of a small matrix that lacks sufficient parallel processing for GPU saturation. This mode minimizes unnecessary memory transmission between the host and the device to improve performance.

For more information on new features and performance improvements, visit the official NVIDIA blog.

Image Source: Shutter Stock