What is the $1 AI chip?
The company selling A$1 chip called AVSdsp, team members from the 24-year-old Shenzhen Shenkeng SQ. CEO Shen Lianjie, who is headquartered in Shenzen Shenzhen PC and consumer graphics chip market, expressed to Lei Fengwang at the first AVSdsp conference. : 'The AIs of other families are big bucks and start off with hundreds of millions of dollars. Compared to those billions of innovations, we have been making image chips for more than ten years. In the past, there were dozens of chips that could be sold in Shenzhen each year. The goal is today. It is a chip that offers a minimum of 1 US dollar. It can be placed next to the master chip in a collaborative IC mode. It can easily and quickly import products with a primary image function.
In this way, we are very clear that, unlike the current mainstream AI chip company's products, the AVSdsp product is a collaboration of the main chip that can help products that have primary imaging capabilities such as home, car, advertising, security, military police, Industry, business, education, etc. to achieve AI functionality.
Shen Lianjie said that advanced academic circles or large research institutions often connect a large number of computers. The cost of training AI for hundreds of thousands of computers is naturally high. Nvidia, Intel and other companies’ products are mostly hundreds to tens of thousands of dollars. For price products, domestic unicorn companies are financing hundreds of millions of U.S. dollars to launch products that compete with Nvidia. They are also powerful and costly. We don’t compete directly with these great companies. We run counter-products and introduce cheap products. AI smart mice.
How does that cheap product do? According to the reporter, AVSdsp has used the tens of millions to buy the 8/16-bit CPU core source code from the foreign small company to 12/24-bit CPU in order to make the image chip earlier. Image algorithm and processor and compile compression engine, and first introduced single-chip network camera in Asia, and later cooperated with universities to develop 32-bit CPU, H.264, face algorithm, cooperation with Shenzhen Institute of Computing and Computing was launched in Shenzhen. Various generations of artificial intelligence chips (AVS01-05) sold 670,000 sets of products several years earlier than Hass and continued to optimize in structure, algorithms, and design to enhance the chip for image processing and smart recognition.
Nowadays, DL-CNN (Deep Learning Convolutional Neural Network) needs to deal with massive amounts of data and huge amounts of computation. The circuit design efficiency will cause a huge gap. 'The concept of deep learning is not difficult to understand, but they are not perfect. Even though the core theory is still like a black box for scientists around the world today, we are all afraid of lagging behind. So we are desperately trying to grab talents, expand, and are desperately working on it to create a variety of architecture transformers. There are many large companies and many departments. , Considering that it will make its structure more and more complex, resulting in poor performance can only be as constant as the mining machine performance and process improvement, the cost is difficult to cheap. ' Shen Lianjie said.
AVSdsp can use tensorFlow or Caffe platform frameworks to build more sparse models at the AI training end, but it is effective in identifying inference ends, saving power, and low cost. It does not consider the design of compatible multi-instruction sets, but adopts an efficient and flexible architecture. Compact, but compatible with the training end, more superimposed into a wide range of high-precision twin networks and enhanced learning to enhance on-the-job training for detection of more than a dozen scenes, plus internal highly efficient data multiple emission and collection accelerators, It can save the operation thread and storage and transportation unit greatly, it is faster than the AI chip already on the market, and the cost is lower (tens to hundred times). And the DL-CNN can learn all kinds of graphics that need to be recognized. Accuracy Also greatly increased, there is no problem in the past that the machine learning algorithm application scenario is not universal.
Therefore, through different image interfaces (Mipi, DVP, YUV, 656, 1120), AVSdsp's AI chip is placed next to the main control chip in the form of a cooperative IC without affecting the master control work, and the image to be recognized will immediately Through the output interface (IIC, UART, Video) to inform the main control chip, to play the role of smart assistant, simple, fast, low-cost import of products that have established primary audio and video processing capabilities, upgrade to artificial intelligence products.
Is 1 dollar a gimmick or a way to survive?
The reporter learned that the $1 AI chip is actually the way for AVSdsp to survive. As mentioned above, AVSdsp's team is based on the technology of Reliance Technology, which is based on the scanners, digital cameras, and MP3 chips introduced in Shenzhen. In 2003, Yiqiang Technology successfully returned to Taiwan for listing. The total chip sales in the past ten years exceeded hundreds of millions of chips. However, the financial turmoil in 2008 triggered internal division of the company. Shen Lianjie and Yiqiang CEO disagreed, so he resigned as general manager. After leaving the post with a number of front-line supervisors, AVSdap was created to specialize in high-end security imaging chips.
However, the reality is very cruel. The high-end processing and image recognition chips they sell sell for $5 and $10 are not very good, and even price cuts are difficult to improve. Instead, they are simple, inexpensive, and easy-to-use algorithmic processing chips. Only eight are sold. Most of Mao’s can sell millions of pieces. It is also a profound understanding that Shenzhen is not the cheapest, only the cheaper and hard truth.
Shen Lianjie said: 'Now that the AI deep learning algorithm has hit the tsunami, no one can stay out of it. It has been inherited for many years and is not as good as early determination to embark on the rush to lead the introduction of the simplest, fastest, and cheaper AI chips. Those great products are left to Nvidia. , Intel, Cambrian, Horizon and other companies to do.
Therefore, we can understand that for AVSdsp, the $1 chip is neither a gimmick nor an attempt to attract investors, but is based on its special experience to learn how to survive.
AI is still a tall technology at present, but Shenzhen AVSdsp has to launch a $1 chip. This price is not a gimmick but is the company's survival. Shen Lianjie also stated that we can do better. Chip, but the price will be higher, but I would like to do is 1 dollar chip. As for developers or AVSdsp customers are concerned about the AI chip product will be launched, he said that we plan to launch DL in 2018 - CNN processor FPGA version, the most simple and fastest low-priced Mipy DL-CNN AI processing chip was introduced at the end of 2018, but the characteristics of small companies are flexible, so when to launch more depends on the customer.
In addition, the field of AI applications is very extensive. Shen Lianjie also stated that this is the first press conference for AVSdsp. The reason why it is necessary to develop a deployment is not only to tell customers that we have a US$1 product, but more importantly, to understand the needs of our customers. In which areas, we have better products. In particular, he pointed out that in the area of AVSdsp's important security, the current market has gradually solidified, and it is difficult to further innovate and develop. Therefore, it needs to find more applications such as automobiles and households. As well as a large volume of fast shipments of toys and other markets.
The reporter believes that since it is only a US$1 collaborative IC, the chip can only help the products that have primary image function to upgrade the limited AI function. AVSdsp also calls it AI smart mouse Mipy, and the performance cannot naturally be compared with the AI chip on the tall one. By the same token, whether or not this is commonly understood by people with different AI chips will also have different insights. However, when many manufacturers have introduced higher-performance products, AVSdsp's team has no financing to sell chips and runs counter to the introduction of simple low prices. The price of the chip helps the existing product to achieve some of the image's AI functionality. Isn't it another idea to participate in the AI competition?