As early as June 20, 2016, China Star Microelectronics pioneered the release of China's first embedded NPU chip - Xingguang Smart One; on May 18, 2018, Zhongxingwei was at the peak of the 2018 Songshan Lake IC innovation. At the forum, the second-generation artificial intelligence chip - Starlight Smart No. 2 was introduced. It is understood that there are tens of thousands of chips introduced by Vimicro in the field of security monitoring. At present, China Star Microchip has shipped more than 300 million US dollars.
After China Star Microelectronics released China's first embedded NPU chip in 2016, after two years of development, the investment and application cases in the AI field showed a blowout situation. So, at what stage is the current artificial intelligence chip? The ultimate form of the future AI chip is Zhang Yiongong, chief technology officer of China Star Micro-Intelligence Co., Ltd. gave the answer, and introduced the micro-second generation artificial intelligence chip of China Star – Star Smart II.
Where is the artificial intelligence chip?
From an industrial development point of view, artificial intelligence will generally go from technology-driven to application-driven to commercial-driven development. At present, the core technology of artificial intelligence has entered the application-driven phase and is on smart phones, security monitoring, and ecology. Protection, smart homes, unmanned systems, aerospace war industry, high-speed railway operation and maintenance systems began to introduce large-scale import.
Zhang Yinong said that artificial intelligence is now in the application-driven stage. The key to application drive is to bring real products to the industry. The current investment in artificial intelligence, the application case presents a blowout situation, will quickly promote the development of artificial intelligence, and in the application Once driven, artificial intelligence will enter the business-driven phase. At that time, AI will be as ubiquitous as the Internet. At this time, AI will evolve into a contest of business models.
In the current application-driven phase, AI applications have two approaches: front-end and cloud. Among them, the cloud applications include new retail smart portraits, carrier intelligent network management, etc.; and in the front-end, AI applications in security monitoring are A good landing scenario. Through front-end smart cloud intelligence, combining the front end with the cloud, you can make real-time online reasoning and data transmission better.
Over the past ten years, security monitoring has gone through analog monitoring, digital monitoring, and network monitoring. Now that it has evolved into intelligent monitoring, it is also an evolutionary process. From the perspective of intelligence and monitoring, the core of the evolution of security monitoring lies in monitoring standards. Coding and ideas, and the introduction of artificial intelligence elements. In the past coding standards introduced artificial intelligence, to some extent the code rate is difficult to decline, and now the use of artificial intelligence deep learning can make the code rate continue to decline.
In addition, from an application-driven perspective, the application of face recognition and vehicle identification is endless, how to effectively unify information and send it to the backend, and also rely on standards. Through the integration of SVAC national standards and specialized information channels, this is also The real place in the industry application.
ASIC chip will be the final form
In recent years, with the continued burgeoning virtual currency market represented by Bitcoin, a number of 'mining' companies have emerged. The most well-known one is Bitland.
The reason that Bitland has emerged in the miner's market is mainly due to its design of ASIC chips for bitcoin mining machines. Because it is more efficient than CPUs and GPUs to use dedicated ASIC chips for mining.
Zhang Yinong believes that as long as the algorithm is mature enough, and there is enough power and power consumption to promote, artificial intelligence chips will soon go to ASIC.
From the perspective of deep learning of artificial intelligence, in the field of blockchain mining, the chip design from 2011 to 2013 has gone through a complete process from CPU → GPU → FPGA → ASIC; and from 2013 to 2018, on the process From 0.13 micron to 7 nm process, the 10 process levels in the middle are also quickly finished, which is much faster than the average chip design company's two-year iterative process. Therefore, Zhang Yinong emphasized that if there is severe demand, there will be rapid In the future, the ASIC must be the mainstream of artificial intelligence chips.
In the design and positioning of the AI chip, there are different thinkings about whether the AI chip is the main processor or the co-processor in the industry.
In fact, doing co-processing is a certain benefit. The chip design of the co-processor is relatively simple and does not require more industry applications. However, Zhang Yinong believes that AI applications are indispensable for DDR and Memory bandwidth. The processor has some limitations on the feasibility and ecological development of large-scale chips in the future. Therefore, the micro-selection of ChinaStar is a single-chip SoC integration solution.
Of course, Zhang Yinong also emphasized that the design of artificial intelligence chips must be tailored. At present, in the application-driven phase, the computing power of the chips does not need to exceed the performance indicators of other functions; instead, it is necessary to pursue a balance between performance and cost, and pursue the use of unit rates. The flexibility and energy consumption ratio.
Third-generation AI chip released next May
Based on the design path of the above artificial intelligence chip and the application-driven stage, how does a standard artificial intelligence application complete the deployment?
Zhang Yinong disclosed that the deployment of a standard artificial intelligence application can be divided into three steps. The first step is to conduct large-scale training in the cloud. Training includes marking, floating-point model, etc. This model will generate a very large parameter set; In the second step, the front-end application must reduce the model parameters. In this process, it involves retraining. The retraining time is much less than the first step. The third step is to download the fixed-point reduction model to the hardware. Simulation optimization and front-end reasoning.
In the specific application deployment, artificial intelligence is used to lock suspects into cases, and the first step is to collect and encode the information. The second step is to identify the extracted objects and identify the trajectory through deep learning. Label the structured information, directly embedded in the standard and upload it in the form of an ASIC. Step 4: Access the public security platform. Follow-up detection directly searches for structured tags and locks in suspects.
This set of standardized artificial intelligence application deployment is exactly the strength of China Star Micro in the field of artificial intelligence. Judging from the current feedback, Starlight Smart, which has been mass-produced by China Star Micro, has unique advantages in the field of video surveillance for identification, classification, and tracking. Compared with traditional solutions, such as false positives and high false negatives, the accuracy of Vimicro's micro-artificial intelligence solution has been greatly improved. The actual occupancy rate in criminal investigation is more than 90%, and there is a huge space for future applications.
Compared to Starlight Smart One, the new second-generation product VC0718P from China Star Microelectronics Co., Ltd. – Starlight Smart II has higher performance, supports the national standard SVAC2.0 video and audio codec, supports up to 30fps@1080P video real-time encoding , built-in high-performance multi-core processor, can run multiple AI algorithms, realize front-end intelligence, support CABAC, bi-directional B-frame, ROI, SVC, monitor special information, encryption and decryption and other SVAC features, support BT.1120 data output, and external AI Coprocessor seamlessly connected.
Among them, the second-generation NPU adopts a distributed architecture and parallel computing method to provide a more flexible hardware solution; The computing capability can meet the real-time classification detection requirements of 1080P@30fps, which is 16 times that of the first-generation NPU. The network simplification technology has greatly reduced the data throughput, further improved the chip's energy consumption ratio and efficiency, greatly increased the on-chip SRAM capacity, reduced the pressure on the memory bandwidth, and greatly improved the overall efficiency compared to the first generation. .
From the definition of artificial intelligence chips in 2014 to the completion of the design architecture in 2015, and the official release of Starlight Smart One in 2016, and the release of Starlight Smart II, the route layout of Vimicro in the field of artificial intelligence has not stopped, it is reported that The micro-third-generation artificial intelligence chip of China Star is also in planning, and will be officially released in May 2019.