1 Introduction

CUDA

CUDA is a general-purpose parallel computing architecture launched by Nvidia in 2006, which aims to solve parallel computing problems on GPUs. Its ease of use and convenience can facilitate developers to program the GPU conveniently, and make full use of the parallel capability of the GPU to greatly improve the performance of the program.

Since the birth of CUDA, the CUDA ecosystem has also developed rapidly, including a large number of software development tools, services and solutions. CUDA Toolkit includes libraries, debugging and optimization tools, compiler and runtime libraries.

ROCm

AMD ROCm is the abbreviation of Radeon Open Compute (platform). It is an open source software development platform developed by AMD in 2015 to benchmark the CUDA ecosystem for HPC and ultra-large-scale GPU computing. ROCm only supports the Linux platform.

Similarly, ROCm includes a series of development tools, software frameworks, libraries, compilation tools, programming models, etc.

The figure below sorts out the comparison between Nvidia's CUDA ecology and AMD's ROCm ecology.

2 Comparison between CUDA ecology and ROCm ecology

2.1 Programming model and API

NVIDIA	AMD	Functional description
CUDA	HIP	Provide a comprehensive environment for C/C++ development of GPU accelerated programs, API, Runtime, compiler, debugging tools, etc.
OpenCL	OpenCL	An open and free standard for general-purpose parallel programming of heterogeneous systems, and a unified programming environment that facilitates software developers to write efficient and lightweight code for high-performance computing servers, desktop computing systems, and handheld devices.
OpenACC		Parallel Computing Instructions, the GPU parallel programming model most commonly used by researchers and technical programmers.
	OpenMP	OpenMP is a specification of a set of compiler directives, library routines, and environment variables that can be used to specify high-level parallelism in Fortran and C/C++ programs.

2.2 Compilation and Toolchain

NVIDIA	AMD	Functional description
NVCC	ROCmCC / HCC	translater
CUDA-GDB	ROCgdb	debug tool
	HIPify	Convert CUDA native code to HIP native c++ code
Nvidia Nsight	ROCm Profiling Tools	performance analysis tool
nvidia-smi	rocm-msi	Tools for System Administration Interface and Command Line Interface

2.3 GPU accelerated library

CUDA和ROCm的基础框架提供众多的支持库，包括基础数学库、AI支持库、通信库、并行库等一些列，下面将列出来做个对照：

数学库

NVIDIA	AMD	功能描述
cuBLAS	rocBLAS	基本线性代数库（basic linear algebra，BLAS）
cuFFT	rocFFT	快速傅里叶变换库（Fast Fourier Transforms）
CUDA Math Library		标准数学函数库
cuRAND		随机数生成（random number generation，RNG)
cuSOLVER	rocSOLVER	密集和稀疏直接求解器
cuSPARSE	rocSPARSE / rocALUTION	稀疏矩阵BLAS
cuTENSOR	rocWMMA	张量线性代数库
AmgX		用于模拟和隐式非结构化方法线性解算器

并行算法库

NVIDIA	AMD	功能描述
Thrust	Parallel STL / rocThrust	C++并行算法和数据结构库

图像和视频库

NVIDIA	AMD	功能描述
nvJPEG		用于JPEG解码的高性能GPU加速库
Nvidia Performance Primitive		提供GPU加速的图像、视频和信号处理功能
Nvidia Video Codec SDK		硬件加速视频编码和解码的一整套API、示例和文档

通信库

NVIDIA	AMD	功能描述
NVSHMEM		OpenSHMEM标准的GPU内存，具有扩展以提高GPU性能。
NCCL	RCCL	多GPU、多节点通信

深度学习/人工智能库

Nvidia

AMD

cuDNN：深度神经网络基元库

TensorRT：用于生产部署的高性能深度学习推理优化器和运行时

Nvidia Riva：用于开发交互式情景AI会话应用的平台

Nvidia DeepStream SDK：用于基于AI的视频理解和多传感器处理的实时流分析工具包

Nvidia DLI：用于解码和增强图像和视频以加速深度学习应用的便携式开源库

MIOpen：AMD的深度学习基元库，提供不同运算符的高度优化和手动调整实现，如卷积、批量归一化、池化、softmax、激活和递归神经网络（RNN）层，用于训练和推理。

MIGraphX：AMD的图形推理引擎，可加速机器学习模型推理。AMD MIGraphX可以通过直接安装二进制文件或从源代码构建来使用。

MIVisionX：MIVisionX工具包是一套全面的计算机视觉和机器智能库、实用程序和应用程序，捆绑在一个工具包中。AMD MIVisionX提供高度优化的Khronos OpenVX™和OpenVX™扩展的开源实现沿着支持ONNX和Khronos NNEF™交换格式的卷积神经网络模型编译器和优化器。