Today, the high-light moments of the Internet have passed, and the era is driven by core technologies. Semiconductors is a market opportunity of hundreds of billions of dollars. It is reported that China's annual import of semiconductor technology, chips, integrated circuits and other aspects exceeds 300 billion yuan, far exceeding the import of other energy and resources.
In the past two years, many domestic players have entered the AI chip. In addition to Huawei and Alibaba, they have also produced such as Horizon Robotics and the Cambrian. A unicorn of the first class. In April of this year, the ZTE crisis broke out, which gave birth to the domestic “core-making” fever, and domestic capital will pay more attention to the semiconductor industry.
In fact, since 2005, Northern Lights venture capital has been in the semiconductor industry. At this year's Aurora Semiconductor Industry Communications Conference, Yang Lei, Managing Director of Northern Lights, gave a speech on "Scenario Definition Computing" to discuss semiconductor technology development and application opportunities, as well as the investment strategy of Aurora Borealis in this field.
The end of Moore's Law, semiconductor investment "time and place"
The development of semiconductor technology is inseparable from Moore's Law and Denard Scaling. The former refers to the electronic components in a semiconductor unit area every 18-24 months. Doubled; the latter means that under Moore's Law, every time the number of transistors is doubled, lower power consumption can be achieved.
In 2010, it was still in Moore's Law. In an effective era, the Northern Lights strategy is “non-Moore's Law Drive”, which is to find technologies that can be independent of Moore's Law.They believe that unique algorithms, special processes, etc. are likely to establish Moore. Opportunities beyond the law.TiDAl System, a NAND flash controller company invested by Northern Lights, was acquired by Micron, the largest storage company in the United States, only two years ago, and received an eight-fold return on investment.
However, in 2015, the development of process technology made the transistor size approach the physical limit. As the transistor size became smaller and the number of transistors per unit area increased rapidly, the static power consumption on the 20nm chip exceeded Dynamic power consumption. Moore's Law has slowed down and investment logic has changed. The Northern Lights realized that the combination of soft and hard may be the next opportunity, and such a company is not a traditional semiconductor company, but more like a system company.
Until now, the annual increase in single-processor core performance has dropped to 3%. 36氪 has interviewed David Patterson, the 2017 Turing Award winner. He judged that Moore's Law has been over for more than 50 years.
In this regard, Yang Lei believes that: Strong> Although the end of Moore's Law has brought enormous challenges to the semiconductor industry, it is a good opportunity for semiconductor investment. First, Moore's Law is over, innovation is the first priority, and games based on capital and scale are no longer good. IBM in the era, Intel in the PC era, and Qualcomm in the mobile Internet era, the emerging era of intelligent Internet of Things also needs to run out of mainstream semiconductor chips,and the vast majority of smart Internet of Things And scenes are the first time in China. The closer you are to the scene, the greater the chance.
For the smart Internet of Things era,The Northern Lights believe that its core trend is decentralization and calculation by scene definition. There are three points for specific splitting:
The application of the intelligent Internet of Things era requires low latency, low power consumption, and high privacy, and is computationally intensive. Therefore, it is necessary to split the cloud and the edge and complete the calculation on the edge node.
The splitting of clouds and edges or the domino effect, with the underlying computing architecture changing, middleware, PaaS, System-level software and databases will change, which will also generate a lot of 2B opportunities.
In the edge computing scenario, future AI chips are likely to be neither TPU nor GPU. Although Google's TPU has low power consumption, it is not flexible and has poor programmability; while the GPU is flexible, it consumes a lot of power. For startups, finding the balance between the two is an opportunity.
Amphitheatre’s “Denlin Technology” starts with balancing flexibility and efficiency, and software-defined heterogeneous computing. There is a software on the computing of a heterogeneous platform.When the upper layer software is changed, it will not affect the underlying hardware.
The smart Internet of Things era will generate 100 times more nodes than the 10 billion devices in the mobile Internet era, most of which It is the end and will grow at a geometric level. Sound-related sensors represented by silicon wheat, CMOS image sensors, LiDAR, millimeter wave radar, X-ray, infrared, and ultraviolet are the most important sensors.
In the future, the end will be simplistic and intelligent. Simplification is reflected in lower cost, lower power consumption, and higher security. Intelligent is reflected in images and visual and sound. For the end intelligence, the battery is a challenge, and the limitation of battery power also limits the computing resources of the smart end. At this time, the combination of soft and hard is particularly important.
In this field, the Northern Lights are laid out from various scenes.It includes wireless RF BlueX with low-power and low-cost IoT, Black Sesame Technology, which is used in the smart driving field, and GMEMS, which is used in smart voice terminal scenarios.
The current CPU and GPU are all von Neumann architecture, the kernel is computation, and the storage is an edge. Device, this is "computation based storage." In the future, this trend may be subverted, and "storage-based computing" will become mainstream.
Specifically, from cloud intelligence to smart end, the integration of chips in the future will become higher and higher, and it is possible that a single chip can achieve the performance of one module today. Connection, sensing, storage, and calculation functions are implemented on one chip.
There are no particularly successful investment cases in this area. The Northern Lights invested in the next-generation storage architecture ReRAM seven years ago, and it takes time to test whether storage-based computing can be established.
Aurora has invested in 15 semiconductor companies by 2018, almost every year, with a cautious investment style. In 2016, a total of three companies were listed (Exhibition, Actions, Zhaoyi Innovation) and two were acquired.
Aurora Borealis’ investment in semiconductors covers SOC from sensors, wireless connectivity, to CPU, scene definition calculations, Store multiple key chains, all of which are known for their R&D core IP. Yang Lei mentioned: In the future, I hope to see more semiconductor companies no longer simply put together some IP, but each family can have a unique core IP.
In addition to the companies mentioned earlier,The core IP companies of Aurora Borealis also include Cista, Radium Intelligence for Sensors, Spreadtrum (Exhibition) for Processors, OURS for Sensors, Yizhi Technology for DSA, and Crossbar for Storage.
At the time of China’s “core heating”, the Northern Lights has two suggestions for semiconductor industry entrepreneurs and investors:
Escape from the low-end trap: China's semiconductor market currently has low-end demand and products, and the middle and high-end companies have a small proportion, while the semiconductor industry can only earn in the middle and high-end areas. money. Therefore, Yang Lei suggested that investors should focus on investing in semiconductor companies with unique core IP research and development capabilities and focusing on large market space, and should not be transferred to low-end traps.
Catch the category differentiation: the end of Moore's Law, the arrival of the intelligent IoT era, which has spawned a lot of scenarios and needs, such as AI training, reasoning, cloud computing , edge computing, intelligent driving, intelligent central control, etc. For large companies, the need to segment the scene is not the focus of attention.This gives the startup a chance. Starting from the differentiation of categories, making the strongest products in the field, occupying the market space is the way to survive.