What impact will the Internet giant have on the semiconductor industry?In recent days, global Internet giants have announced high-profile entry into the semiconductor industry. Baidu, Ali, Microsoft, Google, Facebook, etc. all announced their actions in the chip field, so what impact will these actions of the Internet giant have on the chip industry? This article brings you a detailed analysis.
Ali, Microsoft, Google, Facebook Internet giants have entered the chip
Let's take a look at the actions of the Internet giant in the chip field in recent years.
Google is the Internet company's furthest company in the chip field. In 2015, Google began deploying its own deep learning acceleration chip TPU in its data center for reasoning. Last year, it launched the second generation TPU, which can take into account deep learning reasoning and training, and was also deployed in the mobile phone Pixel 2. The IPU designed by itself is used to accelerate image processing related applications. In addition, Google has also recruited two major divisions in the processor architecture field, David Patterson and John Hennessy, and hopes to have more output in the chip field in the future.
Microsoft is not far behind. In 2014, it released its own ToF sensor chip designed for use in Xbox Kinect 2 at ISSCC. Then in 2016, it released its own HPU coprocessor chip for use in HoloLens. It also released the Azure Sphere IoT platform, which includes the Microsoft Pluton security module used in the MCU chip to ensure IoT data security. The first Azure Sphere chip will be the MT3620 jointly launched by MediaTek and Microsoft, including Microsoft Pluton.
In addition to Google and Microsoft, Facebook has also begun to take action in the chip field, posting ASIC and FPGA design related job postings on the website. Yann LeCun, head of Facebook AI research, also posted on Twitter that this will be a chip design post related to deep learning. According to industry predictions, Facebook is recruiting engineers to develop chips similar to TPU to accelerate server-side deep learning applications.
Amazon Lab26 is also making chips.
In China, Internet giant Ali first announced the start of the development of artificial intelligence acceleration chip Ali-NPU, its performance will be 10 times the current mainstream CPU, GPU architecture AI chip, and manufacturing cost and power consumption is only half, can achieve more than traditional The GPU is 40 times more cost effective. After that, Ali High-profile announced the acquisition of Zhongtian Microsystems, which currently has an instruction set for China's independent intellectual property rights. Ali's move is believed to strengthen its layout in the Internet of Things. Baidu also recently released "Kunlun", which is scheduled to be completed in 2019, facing the cloud data center market.
Revelation 1: The end of Moore's Law is the era of heterogeneous computing, the Internet giant has to customize its own chip
The Internet giants have laid out chips, which means that the original semiconductor industry is undergoing major changes. Initially, the semiconductor industry as a high-tech industry, its development follows Moore's Law, and the size of the process features shrinks every 18 months. As the process progresses, the manufacturing cost of the same chip will be lower because of the increase in the number of transistors per unit area. The area required for the same chip is reduced, and the performance of the chip is increased as the size of the process features shrinks. Therefore, the semiconductor industry is moving under the guidance of Moore's Law to achieve a high-speed development model in which the performance of the periodic performance is reduced. Due to its high technology and high added value, the chip's gross margin is very high, so chip design companies can live very well by selling chips to downstream companies. In the golden age of semiconductor design, the word "chip design" means high technology and high profit margin. The raw material of the chip is sand, but it can sell tens or even hundreds of dollars, so the amount of successful chips Production is like starting a printing machine. In order to increase shipments and improve design reusability, semiconductor companies tend to favor general-purpose platform chips, such as Intel processors and Qualcomm SoCs. However, as Moore's Law gets closer to the physical limit, the amount of capital required to continue to shrink the feature size is growing, and in other words, as the feature size shrinks, the cost of the chip rises rapidly. The cost of the chip includes NRE cost (Non-Recurring Engineering, which refers to the cost of chip design and mask fabrication, which is one-off for a chip) and manufacturing cost (ie, the cost per chip manufacturing). In advanced process processes, NRE costs are very high due to the complexity of the process. For example, the FinFET process often requires the use of double patterning technology, and the number of metal layers can be as many as 15 layers, which makes the mask production very expensive. In addition, the design rules of complex processes are very complex, and engineers need a lot of time to learn, which also increases the cost of NRE. For chips made by advanced processes, the gross margin of each chip is higher than that of chips made with backward processes, but the high NRE cost means that chips made by advanced processes require more sales to achieve real profitability.
As Moore's Law encounters bottlenecks, the cost of improving process performance by process technology is too high. Therefore, more special designs are needed according to the application, and architecture design is used to further improve performance. This is also the idea of “heterogeneous computing”. In the past, semiconductor companies often tended to introduce platform-based products such as Intel's CPU, Qualcomm's Snapdragon, etc., but such a common platform today is unable to meet the performance and energy efficiency requirements of mobile devices, especially in AR/ Emerging applications such as VR and terminal artificial intelligence.
At the same time, we saw the Internet giants such as Facebook, Google, and Ali marching into the hardware field. In fact, this is also the inevitable development of the Internet company's ecological development: Internet companies have been building an ecological rushing traffic portal before, allowing users to use their services more conveniently; this portal used to be from PCs, mobile phones, to today's PCs. Traditional portals such as mobile phones have been unable to satisfy the appetite of the fast-growing Internet giants, so the giants have thought of building their own hardware platforms as an entry point for their services. Such hardware can be AR/VR devices (such as Microsoft's HoloLens, Facebook's Oculus), new consumer electronics devices (such as Google's Clip auto camera), or IoT endpoints that enter smart manufacturing and smart homes.
When the Internet giant begins to enter the hardware market, it will have a huge impact on the semiconductor industry. First, the Internet giant is pursuing hardware to achieve extreme performance to achieve a differentiated user experience to attract users. As mentioned above, Moore's Law currently encounters a bottleneck. The need to pursue the ultimate experience is to take the path of heterogeneous computing of its own custom chips. It is no longer possible to satisfy the Internet giant's demand for hardware by purchasing chips from traditional semiconductor manufacturers. At least in the core chip section. Therefore, Internet giants such as Facebook, Google, and Ali are all active advocates of heterogeneous computing, designing chips for their own hardware layout or plans, or already designing chips. As a result, Internet companies that were originally downstream customers of semiconductor companies do not need to purchase chips from semiconductor companies. Such changes in industrial divisions will cause major changes in the industry. In fact, such upstream and downstream integration trends have been in the mobile phone field for a long time. A few years ago, consumer electronics giants Apple and Huawei have begun to purchase standard chips to design their own SoCs. In the past few years, Xiaomi has gradually started to make its own SoC chips. Once Oppo, Vivo and other mobile phone giants follow suit to start their own core chips, Qualcomm will be under great pressure. The entry of the Internet giant into the chip can also be understood as the story of the mobile phone field has begun to occur in other hardware areas, which will have a profound impact on Qualcomm and other traditional semiconductor companies.
Secondly, the purpose of the Internet giant to manufacture hardware is only to attract users to enter their own ecology to use their own services. The ultimate profit point is not on the hardware but on value-added services. Therefore, the Internet giant can design the chip for its own hardware without regard to cost, and basically does not consider cost-down. In the eyes of traditional semiconductor companies, the cost of a chip is 10 dollars or 8 dollars can be a world of difference, but in the view of Internet companies, such a cost difference is not a problem at all, how to get good performance. On the other hand, once the Internet companies that design their core chips enter the same field, companies that rely on the semiconductor chip's standard chip hardware system are completely uncompetitive, both in terms of price and performance. The Internet giant of the core chip can implement the dimension reduction strike. Once these hardware companies lose their competitiveness, the living space of semiconductor companies that rely on these customers will be further compressed.
All in all, the Internet giant entered the chip field, first of all in performance considerations, no longer sourcing core chips from semiconductor companies, which impacted the traditional industry division of labor, making traditional chip companies lose a large class of customers; on the other hand, the Internet giant's ecological style of play can Let the self-developed hardware chips not consider the cost, which in turn impacts the traditional hardware companies that purchase chips from semiconductor companies, thereby further compressing the semiconductor company's market. Under these two roles, the traditional business model of semiconductor chip companies must change in order to catch up with new trends.
Revelation 2: Semiconductor industry division of labor shuffling, three roads to deal with Internet business barbarians
At present, the division of labor in the field of semiconductor chips can be divided into several parts such as definition, design, design and finalization. Definition refers to the definition of the top-level architecture and indicators of the chip; design refers to the specific circuit design, including digital RTL design, analog/mixed-signal circuit design, etc.; design finalization means that the chip design is finalized and enters Fab to decide to start manufacturing lithography mask. Membrane means that it takes a mask overhead; manufacturing is done at the Fab and finally the chip is made according to the mask.
Today's division of the semiconductor industry, the most well-known is the Fabless model, that is, the chip design company is responsible for the definition, design and design of the finalization, while the manufacturing is completed in the foundry Fab; Samsung, Intel and other semiconductor giants will also include Define IDM mode to all aspects of manufacturing; in addition, many companies are mixed Fabless and IDM, that is, some products go the Fabless mode, and some products go IDM mode, such as Broadcom's CMOS chip product line is Fabless, while in RF The front-end module product line is the IDM mode with its own Fab. In today's semiconductor industry division, design service companies provide module design IP and assist some companies in design, mainly to assist.
After the Internet giant entered the semiconductor industry, a new model emerged, that is, the Internet company is responsible for defining chips, completing a small part of the design, and spending money to complete the design finalization (Internet companies are not bad money!), the design service company is responsible Most designs, and the foundry is responsible for chip manufacturing. This new mode can be called the Designless-Fabless mode. Historically, semiconductor companies have moved from the traditional IDM to the Fabless model, mainly because the Fab overhead has become a burden for semiconductor companies, and the foundry offers a very flexible option. Today, Internet companies have entered the Designless-Fabless model after entering semiconductors, and outsourcing a large number of designs is mainly due to time costs. The Internet giant is doing chips, and in addition to the ultimate performance, there is also a fast time to market. For them, it would be catastrophic to miss the Christmas/double eleven period. If you want to build your own front-end + back-end design team from scratch like a traditional semiconductor company, and accumulate module IP from scratch, I am afraid that the first chip will be available several years later. This rhythm is not keeping up with the fast iterative rhythm of Internet companies.
So how to achieve high performance and fast listing? The best solution is that these giants recruit their own chip architecture design team to make chip definitions, and use experienced veterans to customize the architecture to meet performance requirements. Specific implementations, including physical layout design and even front-end circuit design, can be handed over to Design service company to do it. An important feature of semiconductor chips is that the details are very important. A small detail such as ESD, heat dissipation, and IR Drop may cause the performance of the chip to be greatly reduced and cannot meet the demand. Therefore, if the specific design work is handed over to an experienced design service company, the risk of detail errors can be greatly reduced, thereby reducing the risk that the chip needs to be redesigned to delay the time to market.
In fact, we have seen in some chips that the Internet giant is responsible for the chip architecture + design services company to provide IP / design services. One example is Microsoft's use of HPU in HoloLens, where Microsoft designed a multi-core architecture to accelerate AR computing, and each core uses Cadence's Tensilica IP.
Microsoft's HPU, heavily used Cadence's Tensilica IP
Internet companies that don't have money are entering the semiconductor industry in an ecological way, like a barbarian invasion. For Fabless' most troublesome design, the cost of filming is not a bad thing for Internet companies. So how do semiconductor companies deal with it? It should be said that there are at least three methods, namely, transforming into a design service company, doing a full stack company, and doing IDM.
The idea of transforming into a design service company is very simple. Since semiconductor companies can't beat the money of Internet companies, then we should be friends. Designing the high cost and high risk of finalizing the film is the pain of the current semiconductor company, then I will not stream, but help the Internet company to do design services, so my risk and overhead is much smaller. On the other hand, relying on the accumulated experience of semiconductor companies in the field of semiconductor design to help Internet companies to design services, it can significantly reduce the time and quality risks required for the listing of Internet-specific chips. As a result, everyone is happy and realizes the optimal allocation of resources. Many semiconductor companies have also noticed the trend of design services and began to move closer to design services. The high-profile public design service business of MediaTek was an important sign of the semiconductor company's turn. MediaTek is a typical company that makes profits by cost-effective operation. It relies on the player to make a comeback and glory in the era of the cottage. However, in the case of Moore's Law, the bottleneck semiconductor company is oligopolistic, its profit margin declines, on the one hand, Dare to use advanced process, worry about the risk is too high, on the other hand, do not use advanced process core