With the recent success of deep learning in the field of artificial intelligence, from CPUs to GPUs, to various application-specific processors, we are embracing another golden age of computer architecture evolution. However, the research and development cycle of processor chips ranges from a few years to a dozen years. To be successful in this new trend of the computer industry, the most important aspects include the perseverance and spirit of the craftsman.

When Huawei invited me to write a reference book for their Ascend AI processors, I accepted their request without any hesitation, due to the high recognition of Huawei as a hardware technology company and the sincerity of its employees in ensuring they are at the forefront of revolutionary technologies.

Huawei’s launch of the AI processor for artificial intelligence computing offers much more computing power and significantly less power consumption. It brings greater possibilities for a wide range of deep learning applications. A famous Chinese proverb says “the journey of a thousand miles begins with a single step.” The Ascend AI processor still has a long way to go. For a high-end processor, the development of the ecosystem and the acceptance by program developers play very important roles, which is critical to the success of this product. The purpose of writing this book is to primarily demystify the Ascend AI processor to the world through exploring its inner design concept. From both the software and hardware perspective, this text introduces good development practices for the Ascend AI processor platforms. Rome was not built in one day. If creating an ecosystem is a journey of a thousand miles, this book is the first step in this journey.

This book can be used as a selective textbook in the field of artificial intelligence processor design. Since the target readers are mainly engineers, we have limited the complexity of abstract formulas, theorems, and theoretical derivations as much as possible. The reader only needs to possess basic mathematical knowledge and programming skills, and there are no course prerequisites. The ideal audience for this book is R&D personnel who require large-scale deep learning calculations in the field of artificial intelligence, computer science, electronics, biomedicine, physics, chemistry, or financial statistics. The book, therefore, provides a comprehensive and effective reference to companies and developers of AI processors.

This book is divided into six chapters. The content includes basic theory of neural networks, computer processor design, open-source neural network architectures, Ascend AI processor hardware and software architecture, programming theory and methods, as well as use-case presentations, etc. By introducing both the theory and practical experiences, we hope to assist readers in understanding the DaVinci architecture used by the AI processor and master its specific programming and usage methods. Through this process, readers will be capable of creating their own artificial intelligence applications.

Many thanks to Xiaolei Wang, Zishan Jiang, and Xing Li for their great contributions to the writing of this book. They have injected great energy into sorting references, text editing, and testing of all the sample code in this book. Without their full commitment, it would not have been possible to complete this book successfully.

Thanks to Ziyuan Chen and several others for carefully editing and modifying the illustrations in this book as well as making the content clearer. After their feedback, the interpretation of images and concepts is more specific and clear.

Thanks to Huawei for the resources and support provided during the writing of this book.

Thanks to the support of Mr. Dongliang Sheng and Ms. Zhifang Zhong from Tsinghua University Press for their careful and meticulous efforts. Their involvement has guaranteed the quality of this book.

There are inevitable omissions and errors in the book, and readers are recommended to provide feedback, as appropriate.