On March 10th, Beijing Skyline Robot Technology R & D Co., Ltd. jointly led by Nanjing Core Semiconductor Technology Co., Ltd., China Intelligent Network Auto Industry Innovation Union (Caicv) “Intelligent Network Auto Vision Perception Calculation Chip Technology Requirements and Test Methods “CSAE standard has passed the provisions of the” China Automotive Engineering Society Standard (CSAE) Revision Management Measures to conduct review, officially included in the 2021 National Standard Development Program (drafting task number: 2021-15).
After several core technologies have been in the past 2021, the intelligent network automotive industry accelerates the stage of entering mass. Recently, during the “two sessions” in the country, the recommendations on automatic driving and intelligent networks also focused on the floor of technology. The standard first is the consensus of the development of the intelligent network, and the “two sessions” period has established a unified technical standard around the test, verification, mass production, etc., has become a focus on topics.
With the “deficiencies” incidents of the automotive industry, there are four national NPC representatives to promote the industrialization of automotive chips during the “two sessions”, including standard development. Among them, the National People’s Congress, the party secretary of the Qirui Auto Co., Ltd., and Chairman Yin Youpi proposed to establish a chip innovation development platform, from standards, standards, talents, technical levels to the chip industry, parts industry and vehicle to support.
The development of my country’s automotive chip industry is far behind the international level, and the domestic chip market share is less than 10%. The car grade chip is critical to the development of my country’s intelligent network car. The horizon of this lecture is “Intelligent Network Car Vision” Perceived computing chip technology requirements and test methods, compliant, adapting to the intelligent transformation trend of the automotive industry.
Visual sensing computing is one of the car intelligent cornerstones. The current intelligent network car’s visual sense computing chip technology has developed very fast, related calculation framework, perceptual algorithm, sensor program has great progress, thereby birth Chip products and visual perception solutions, but how to measure the performance of visual sense computing into new challenges. The industry urgently needs to develop unified evaluation criteria to objectively measuring the product performance of industry participants, which greatly reduces the work of repeated evaluation of downstream manufacturers and accelerates the industry’s survival of the fittest. Standards are greatly reflected in the development of the industry development, in the history of the development of the PC, and the development of the communications industry.
In response to this visual sensing calculation chip (hereinafter referred to as “chip”) AI performance evaluation method, there are two kinds of chip evaluation standards commonly used in the industry: one is peak intensity, but peak integration only reflects the biggest calculation capacity of chip theory. However, there is a big limiter in the actual AI application scenario; the other is the current industry more well-known benchmark tissue MLPERF, the model is small and the speed is lagged behind the evolution of algorithm, and cannot be reflected in time The improvement of the efficiency of the algorithm and the calculation speed of the chip can be achieved under various precision, so the whole picture of the chip Ai performance cannot be described.
Multiple industry organizations at home and abroad are launching related chip Ai performance evaluation criteria, including Mlperf, Zurich Institute of Technology Ai Benchmark, etc., the characteristics of such evaluations are mostly the category of reference testing, depending on the number and update of the model. speed. Due to the evolutionary speed of the chip algorithm, it has caused a phenomenon between the method of evaluating the method of assessing the performance of chip Ai and the development of algorithm. The intelligent network automotive industry is still lacking with the times and can effectively assess the standard of chip Ai performance.
In response to the problem of current chip performance evaluation, the horizon proposes the MAPS (Mean Accuracy-Guaranteed Processing Speed, the Precision Holding Average Frame Rate) Evaluation Method, which is the characteristics of application scenarios, with the premise of accuracy, and to accommodate all of the algorithm Select, evaluate the average processing speed of the chip to the data to provide a new perspective for evaluating the real performance of the chip.
MAPS evaluation method
Based on the understanding of the application of the intelligent network car visual sensing scene, the MAPS evaluation method proposed by the horizon helps to solve the current status of the performance evaluation method of automotive visual sensing computing chip performance. The MAPS evaluation method is intuitive, quantifies the visual sensing computing chip’s real performance, providing clear evaluation results, traction the chip performance optimization direction for real scenes, helps the industry more fully understands the visual sensibility of each chip, find the most suitable Visual sense of knowledge.
Under the leadership of the National Automotive Standardization Technology Commission, China Automotive Engineering Society, China Intelligent Net Auto Industry Innovation Alliance, my country’s intelligent network car key standards accelerated the establishment and promotion, and the standard coverage that has already been established. Car control operating system, C-V2X car network, intelligent network automotive data security sharing and scene data collection, etc., lay the foundation for the large-scale commercial landing of the industry.
The skyline is the edge AI chip leader, which has long been committed to the hardware and software development of AI chips and business landing. The standard “Intelligent Network Auto Vision Sewage Calculation Chip Technology Requirements and Test Method” Standard by Skyline and Core Chi Technology, the United Nations, Guanzhi, Huawei, Guo Zheng Zhi, FAW, De Saiwei, Ideal Car, Baidu The advantages of Industry, China Auto Center Academy, Electronic Four Courtes, etc., the draft is expected to be released within the year will provide test benchmarks to the industry, greatly drive the application and development of automotive calculation chips in the industry chain. In addition to this standard, another standard of the auxiliary area “Auxiliary driving to visually perceived performance evaluation methods and requirements” have also been formally established, helping to improve the security quality and synergistic innovation efficiency of intelligent network car assist driving visual perception . In the future, the skyline will continue to work with relevant standardization organizations, industry associations and industrial chain. The self-controllable and development of the automotive smart chip industry chain will be promoted.