NVIDIA's DALI 0.25 Deep Learning Library Adds AArch64 SBSA, Performance Improvements

DALI is the project at NVIDIA focused on GPU-accelerated data augmentation and image loading along with other tasks while being optimized for deep learning workflows. DALI 0.25 was released on Friday as the latest step forward for this open-source NVIDIA project.

Yes, DALI (short for the NVIDIA Data Loading Library) is one of NVIDIA's open-source projects. DALI currently describes itself as "a library containing both highly optimized building blocks and an execution engine for data pre-processing in deep learning applications."

With DALI 0.25 there is now support for the AArch64 Server Base System Architecture (SBSA) with CUDA 11, new operators, a rework to the resize operator, extensions to their C API, and performance improvements. Some of the performance work includes for arithmetic operations on the GPU and priorities in the CPU thread pool to allow better load balancing of uneven samples. There are also many other bug fixes and other minor improvements throughout. DALI 0.25 comes with full support for the official build of CUDA 11.

More details and download links for DALI 0.25 via GitHub. DALI has been developed as open-source the past two years now and is out under the Apache 2.0 license.