NVIDIA for HPC

原文链接

DALI: 将图像数据的预处理，从CPU挪到GPU上来做，从而加速之；预处理和训练网络overlap起来；

Libraries：

An easy replacement for CPU libraries.

Standard Languages:

NVIDIA C++17编译器，add support for execution policies on the standard template library (STL)？？？

Directives：

Directive-based programming models provide an easy on-ramp to parallel computing on GPUs, CPUs, and other devices.

If standard languages don’t have the flexibility or features you need to get good performance, augment with directives and remain portable to other compilers and platforms.

CUDA：

CUDA is a parallel computing platform and programming model designed to deliver the most flexibility and performance for GPU-accelerated applications. 

To maximize performance and flexibility, get the most out of the GPU hardware by coding directly in CUDA C/C++ or CUDA Fortran.