Essence Sharing 丨 An Algorithm Framework Is the Link Between AI Chips and Commercial Values
Technical articles
Released on2022-06-20 17:53:56
Recently, the first session of the ICPA Series Forum with the theme of “Intelligent Computing Ecology for Joint Creation and Mutual Win” ·was successfully held.

Recently, the first session of the ICPA Series Forum with the theme of “Intelligent Computing Ecology for Joint Creation and Mutual Win” was successfully held.
Mr. Sun Yue, Deputy Director of the Artificial Intelligence Development Division of the Shanghai Municipal Economic and Information Commission, and Mr. Yang Fan, Chairman of ICPA, Co-founder of SenseTime and Vice President of SenseTime Group, delivered speeches. Prof. Lin Dahua, Co-founder of SenseTime and Chief Scientist of SenseCore, Mr. Zhou Bin, CTO of Huawei’s Ascend Computing Business, Mr. Lu Tao, President of Graphcore’s Greater China Region and Global CRO, Mr. Zhang Yalin, Founder and COO of Enflame, and Mr. Liu Bo, Deputy General Manager of SSCT, delivered keynote speeches to discuss the win-win way to accelerate the implementation of AI computing power industrialization and the construction of algorithm ecology.
Prof. Lin Dahua delivered an exciting speech themedAlgorithm Framework: The Link Between AI Chips and Commercial Values. He said: “The development of AI chips and the computing industry derived from them is not a vision that any enterprise can achieve alone. Rather, it require that both upstream and downstream institutions and manufacturers with different roles jointly build a prosperous ecosystem. Under such a system, SenseTime plays a very important role. How to coordinate the upstream and downstream of technology is also a topic we all face together.”
This article is intended to comb the speech content of Prof. Lin Dahua.

An algorithm is the key bridge to connect application values and chip computing power.
The core of artificial intelligence and intelligent computing is the AI computing chip. Whether it is a general-purpose graphics processor represented by GPU or AI-specific computing chip that has emerged in recent years, they all constitute the core of the entire AI computing power.
For users, the application of AI technology in various vertical fields such as autonomous driving, smart commerce, smart city, and smart healthcare has truly brought the value of artificial intelligence to production, life, and work.

The wide application in different industries is supported by a series of key AI algorithms, such as classification, detection, segmentation, and other algorithms in the field of computer vision.
These algorithms can be divided into several small categories, but in each category, their iteration and evolution speed is very fast and there are many types. At present, SenseTime has produced more than 30,000 algorithm models for different industry applications.
These algorithm models pose great challenges to hardware adaptation. The algorithm adaptation of each model requires a lot of work, and different applications also involve different algorithms. Algorithms are the key bridge connecting application value and chip computing power.
In the field of artificial intelligence, algorithm researchers do not directly write algorithms on the underlying API provided by a chip, because most researchers and algorithm engineers do not understand the architecture of the chip, nor do they need to understand it in their specialties.
Therefore, in the middle of connecting algorithm researchers and underlying computing infrastructure, a series of basic software systems of deep learning frameworks have been derived, such as TensorFlow, PyTorch, MindSpore, and SenseParrots self-developed by SenseTime. Each training framework plays a different role in the industry and has embarked on its own differentiated development path.
There are many training frameworks between algorithms and chips, and different frameworks come from different institutions and enterprises, so that no fixed interface is formed. As a result, AI training chips have to not only adapt to different frameworks, but also support diverse algorithms, while the diversity of algorithm framework interfaces leads to a sharp increase in the workload to support algorithms by AI chips and therefore brings high adaptation costs, which become obstacles to the rapid iteration and market entry of AI training chips.

Build an efficient and prosperous AI ecology with open-source algorithm systems
To build an ecosystem that can help chip development, the most important thing is to break down existing obstacles and bottlenecks, so as to promote the upstream and downstream common development and prosperity of the entire ecosystem.
How can we achieve this goal? Algorithms directly support applications and, from scene classification and license plate recognition algorithms in the field of smart city to semantic segmentation and lesion detection algorithms in the field of medical image analysis, usually go through a decomposition process based on application scenarios, finally forming a big algorithm family tree.
After years of development, SenseTime has formed a deep accumulation in AI algorithms, but only by using these algorithms for the entire industry can it guide the development of both upstream and downstream sections of the entire industry and maximize their values. Based on the understanding of algorithm decomposition from the perspective of applications, SenseTime launched the OpenMMLab open-source algorithm system in 2018.
In the recent 4 years, OpenMMLab has developed a bigger and bigger international influence,and already won 58,000 Stars on GitHub, which outnumber those of PyTorch, the top-class deep learning framework in the industry. At the same time, OpenMMLab has also supported the publication of thousands of papers and assisted dozens of contestants to win championships in vertical fields. In addition, it has made achievements in terms of ecological influence.

OpenMMLab is also widely used in the commercial field.At present, there are more than 600 enterprises and scientific research institutions using OpenMMLab to perform technical R&D, and many of them are large State-owned Key Enterprises, head technology enterprises and Internet enterprises, marking that OpenMMLab has formed a wide influence.
Based on the open-source algorithm system, SenseTime hopes to realize the two-way value guidance involving both training frameworks and AI chips, and join hands with industry peers to build an efficient upstream and downstream ecosystem of artificial intelligence.
Build standardized systems, and create technical infrastructures for the synergistic development of algorithms and chips
In order to promote the collaborative development of training frameworks and AI chips, SenseTime decomposes algorithms into various operators, and informs developers which operators are the most important in the entire algorithm layer and application scenario layer. Such guidance can help industry partners to apply limited computing power resources to truly valuable business scenarios.
At the technical level, SenseTime has built two standardized systems, namely the algorithm classification system and the standard operator interface system.
For the construction of the algorithm grading system,SenseTime has initially divided algorithms into the following three levels based on community feedback from multiple dimensions such as influence, performance, and deployment breadth: P0, P1, and P2:

• Level P0 refers to an algorithm that must be fully supported by any chip;

• Level P1 refers to an algorithm that widely serves in business scenarios but is not absolutely necessary;

• Level P2 refers to an algorithm with relatively less usage and attention.

On the basis of algorithm grading, SenseTime follows the principle of “business application as orientation, algorithm as path”, and has formed a complete set of adaptation and evaluation systems for training frameworks, inference engines, training chips and cluster environments, thus giving downstream software and chip manufacturers a very definite and clear optimization and adaptation guide.

For the standard operator interface system, SenseTime extracts standardized operators according to the algorithm guidance. In this process, SenseTime has completed two important tasks:

• The first one is to unify operator interfaces and function signatures, including operator interface and input/output information;

• The second one is to create a consistency test suite, including standardized test cases and related tool systems, which can verify the correctness of operators and evaluate the execution efficiency in different environments and configurations.

These two tasks provide a standardized path for evaluating different chips, and the whole set of technical infrastructure can promote industrial collaboration.

In addition, today’s adaptation methods also have obvious advantages over traditional adaptation methods. Traditionally, chips and frameworks follow the one-to-one adaptation method, but it is difficult to reuse its experience across chips or across scenarios. SenseTime’s many-to-many chip-frame adaptation process is based on the standard operator interface system, and can minimize communication costs, work difficulty, and adaptation workload. Hundreds of standard operator interfaces can be connected at one time, and after passing the conformance test, can be automatically adapted to different algorithms and chips to enable faster iterations.

In the future, SenseTime will work to build more AI technologies, thus actively promoting the upstream-downstream cooperation and coordination of the industry and providing support for more efficiently opening the closed chip-to-value loop.