First, the evolution of the chip industry
The overall development of the chip industry began in the 1960s, and it was exponentially developed at first. So with Moore's Law, which doubled the integration every 18 months, it can be said that the development at that time was very fast. The logic behind Moore's Law is that with the evolution of process technology, the manufacturing cost of the same chip will be lower, the number of transistors per unit area will increase, and the area required for the same chip will shrink. If the process speed is too slow, then This means that the cost of chip production remains high, resulting in an inability to expand profits. However, if you are desperate to use all of your capital to develop a new process, the risk is too great, and once the R&D fails, the company is finished.
Moore found that the process evolution of successful semiconductor manufacturers in the market was about double the number of transistors integrated on semiconductor chips every year, so he wrote a famous paper to tell everyone that this speed of development is a good compromise between cost and risk. The future development of the industry can follow this speed.
It can be said that the ultimate driving force behind Moore's Law is actually economic factors. The positive impact it brings to the market is that with the evolution of the semiconductor process, the performance of the chip grows exponentially, which drives the performance of electronic products. Development, the electronic market is full of vitality. In the first three decades of Moore's Law, the development of new process processes is not difficult, but as transistors become smaller and closer to the boundaries of macrophysics and quantum physics, the development of advanced process processes is becoming more and more difficult. Research and development costs are also getting higher and higher. If the process continues to shrink the feature size exponentially as stated by Moore's Law, there are two obstacles, one is economics and the other is physics.
The obstacle to economics is that as the size of features shrinks, the cost of chips rises rapidly. This cost includes NRE costs (Non-Recurring Engineering, which refers to chip design and mask fabrication costs, which are one-time for a chip) and manufacturing costs (ie, the cost per chip manufacturing). Some people have calculated that with the latest technology, the NRE of a chip will be more than 10 million US dollars. Such a high NRE means that very, very high chip production is required to dilute this cost. In other words, if the chip's production is not enough, then you will not be able to use the latest process, only the older process. This partially broke the logic of Moore's Law, "Investment Development Process - Reduced Production Cost of Chips - Continued Investment Development Process with Partial Profits".
The obstacles in physics are mainly derived from quantum effects and lithography precision. When the transistor is too small, it will encounter various problems. For example, when the feature size is reduced to 10 nm, the thickness of the gate oxide layer is only as thick as ten atoms. At this time, many quantum effects are generated, resulting in transistor characteristics. hard to control.
I believe that the development of the chip industry has reached a bottleneck period, and there are three development strategies, namely More Moore, More than Moore and Beyond Moore. The first type of More Moore, meaning to continue to follow the old path of Moore's Law, continue to reduce the size of the transistor; the second More than Moore, meaning that the performance of the chip system is not simply reduced by the transistor, but more Ground circuit design and system algorithm optimization. Secondly, the improvement of integration does not have to rely on putting more modules on the same chip, but it can be realized by packaging technology, such as Intel's EMIB, TSMC's InFO, etc. (Apple's processor uses InFO technology) The third Beyond Moore means simply developing new devices without CMOS devices, but this seems to be far away. The future estimate is that More Moore will be combined with More than Moore, and more than Moore will become more and more important over time.
In fact, I am still very pessimistic about the future. If there is no new device in this bottleneck period, it should last for 5 years and 10 years or even longer. Just like the steel industry, it has already entered a gradual period, and there is no special new technology. If the industry wants to have a relatively large development, then unless new materials are found to replace steel. The same is true in the semiconductor industry. I don’t know when the new device is coming out. Sometimes the scientific thing is very accidental. It may suddenly find that a certain material can be used, but if you can't find new materials, There is no way to make a big breakthrough.
Second, NV/Google/Intel/AMD's different strategies in AI chips
In fact, artificial intelligence computing is still divided into many areas, chip applications have about two extremes: one is the high-power and high-computing chip for the cloud server, the high-performance supercomputer (HPC) path; the other is For artificial intelligence chips in terminals (such as mobile phones), this pays special attention to low power consumption, and the requirements for computing power are not particularly high.
In the field of cloud server, Nvidia's GPU has become an indispensable part of the server because of the huge amount of data to be processed, but Nvidia has not yet planned to do it on its own, so in the cloud market of artificial intelligence, Nvidia provides Hardware rather than platform. In the field of terminals, Nvidia's GPU-based artificial intelligence platform consumes too much power on the one hand, and too high computing power on the other hand leads to excessive cost, so it cannot compete with custom chips. In fact, the most advantageous application scenario of Nvidia's artificial intelligence platform is in the middle of the above two situations, that is, the amount of data is medium, the computing power is still relatively high, and there are certain requirements for power consumption but not harsh, such as the ADAS market. . Nvidia's artificial intelligence platform is perfect for both computing power (10-100TOPS) and power consumption (10-100W), so it's not surprising that it dominates the autonomous driving market. It was also released at CES in January 2017. Autonomous driving related products.
Intel's words, from three aspects. First, in the cloud market, it is the biggest player, and is actively preparing to compete with Nvidia, because Intel's business in HPC is inherently familiar, and Nvidia has not been into HPC for a few years, only the new player in this market. It was probably in 2013 that people discovered that the original GPU could be used for deep learning, and had never known this before. Speaking back to Intel, it launched a dedicated FPGA-based deep learning acceleration card after the acquisition of Altera, which can be used in the cloud. In addition, after Intel acquired Nervana, it is actively promoting the Knight Mill Xeon processor optimized for its AI with its technology, and the goal is also in the cloud. Second, on the vehicle side, Intel and Mobileye and BMW formed an autonomous driving alliance, Mobileye provides sensor chips and algorithms, Intel provides a cloud computing platform, and BMW provides cars. Third, on the mobile side, Intel acquired Movidius, but has not seen big moves. Therefore, I expect that the artificial intelligence chip on the mobile side, if any, will be advantageous to vendors such as Qualcomm.
Let's talk about Google, the chip TPU that it introduced is mainly for personal use. This is a bit like IBM, IBM's first Power PC series chip is also for their own server. So Google is also a similar idea, its chip is not intended to be used by others, in other words, it does not really intend to enter the chip market, and compete with others.
Finally, AMD, its GPU and CPU technology are in the position of chaser, relatively low-key in AI, when the CES announced new products, they did not take the initiative to mention artificial intelligence. The most recent news is probably to cooperate with Alibaba cloud service as a test. AMD's overall idea is still to stabilize, do not deliberately fight with Nvidia who is the first person, it will wait for you to make these things first, then you will be very practical to do the graphics technology. In fact, GPU is inherently in line with the requirements of deep learning. As long as AMD takes the computing power of its own chip, it can quickly enter the field of artificial intelligence.
Third, the opportunities and challenges facing Nvidia
2016 is a year of artificial intelligence explosion. With this shareholder wind, Nvidia's share price has more than tripled last year, which is amazing. At present, Nvidia has a lot of room for technology growth, because Nvidia is transforming into a platform company rather than a hardware company. The GPU will be its core but not all. It needs to be a platform and an ecosystem around the GPU. Development tools such as development platforms, developer communities, and envelope programming languages ​​are also important to support the various components of the GPU. For example, in the notebook PC market, in fact, ARM's processor performance can fully compete with Intel, but why is there basically no ARM processor for notebook computers? It is because ARM does not have any ecology on the notebook PC. Once the platform and ecology are done, even if its technology development is not as strong as it used to be, I believe Nvidia's business value can still be guaranteed. If AMD makes a GPU with the same performance as Nvidia tomorrow, it will not be able to replace Nvidia for a while, because Nvidia has its own development tools such as CUDA and CUDNN.
The biggest risk Nvidia may face is that its current share price is entirely supported by artificial intelligence, but the application of artificial intelligence will not be as suspicious as investors think. In fact, it is very obvious that the application of artificial intelligence now has a big bubble, and everyone expects it to rise in a year or two. But if it doesn't get up in a year or two or if some applications don't really land, then investors may have some backlash. Now it is an overshoot. After discovering that it has not reached the expected level, there will be an undershoot. After several shocks, it will slowly return to the rational valuation.
Fourth, FPGA players and opportunities for startups
FPGA is called "Field Programmable Gate Array". The basic principle is to integrate a large number of digital circuit basic gates and memory in the FPGA chip. Users can define these gates by burning in the FPGA configuration file. The connection between the memories. This burn-in is not one-off, that is, the user can configure the FPGA as a microcontroller MCU today, and the configuration file can be edited tomorrow to configure the same FPGA as an audio codec. Now the main opportunity related to FPGA and artificial intelligence is the configurable operation of cloud server. Currently, there are two players in this field, Xilinx and Altera, which have been acquired by Intel.
One of the most critical issues encountered by FPGAs today is the developer ecosystem. Traditional CPUs are also good GPUs, programming is relatively easy, the language is C++, Java, everyone is familiar with, and has formed a mature system, development environment, ecosystem and developer community are very it is good. But FPGA development usually requires the use of hardware description languages ​​such as Verilog, VHDL, etc., which requires a lot of time for the programmer to master. In this case, the ecology of the FPGA cannot be developed, because the high threshold means that fewer people are done, and fewer people do it means lower visibility and fewer related projects, which in turn leads to the inability to attract developers to participate in the project, thus forming a malignant cycle.
In response to this situation, Xilinx has released a more ecologically improved thing called the reconfigurable acceleration stack, which will use a partial reconfiguration scheme for the cloud. What does that mean? Usually the FPGA configuration process includes the synthesis of the hardware description language, place and route, and finally the bitstream file is generated and written to complete the configuration. In this process, synthesis and place and route take a very long time, up to several hours, and the final bitstream file write and configuration can be completed in one second. FPGA solutions for the cloud In order to achieve fast application switching, it is expected that hard IP (the bit stream accelerated for an application hardware) will be used and written quickly when the application is needed. In the future, the cloud FPGA ecosystem is expected to include not only Xilinx, but also many third-party IP providers, and finally form an App Store-like form that allows users to easily purchase the corresponding hardware acceleration solution and load/switch in real time.
At present, the level of domestic FPGA is weak, and FPGA requires a complete set of software-to-hardware processes, which requires deep accumulation. It is not the best direction for startup companies, and is more suitable for national research projects. But startups use FPGAs, FPGA solutions, FPGA-based development or IP on FPGAs, which is a good opportunity. In other words, you don't go to the phone, but do the app. In fact, there are already many companies in this direction in China, and I know that there are more famous companies.
Fifth, the correct posture of playing artificial intelligence software
At present, the biggest problem of artificial intelligence software is how to implement the technology to solve the needs of consumers. Some algorithms are very technical, but it is not easy to land, such as image classification. I think the better software that I am doing now is the speech recognition/interaction class. The typical one is the Xunfei input method. I was shown a show at the launch of the hammer mobile phone a while ago.
The artificial intelligence algorithm of the image class is currently the hottest direction, and it is much hotter than the speech class. However, it seems that most of the image algorithms can only be used as part of a large system. For example, a security system, the image algorithm can be used to detect whether a person has a knife, but the software can not be established if the software is pulled out separately. Of course, there are also separate software, such as Prisma, the software that was very popular on Instagram is to use deep learning to do image style transformation. To sum up, the artificial intelligence-related software is of course a very good entrepreneurial direction. It is only to find out the innovative selling point. It is not very useful to have good technology.
In the field of artificial intelligence, there have also been a number of good companies in China, and I am concerned. In the field of image detection/face recognition, there are three leaders, namely, Image Technology, Face++, and Shangtang Technology, which provide face recognition solutions for banks and some security systems. In the medical field, there are many companies doing artificial intelligence to help humans determine the disease. A picture is entering this field. In the field of automatic driving, the software is relatively famous for Tucson and the horizon. In fact, the horizon stalls are relatively large, hardware and software are done, in addition to assisted driving, but also with the United States to do smart home. There are also some relatively small areas, such as Jian Huang, which is the map. Hardware, more famous is the Cambrian.
Sixth, the future of the chip industry, in addition to artificial intelligence, there are...
Finally, return to the development of the entire chip industry. Some people ask, is there a law in the chip industry that has a strong and strong, and how likely it is to have a black horse. I think this is the case: in the place of the chip, it will change the application direction every other period. In the 1990s, the most popular is the multimedia computer, which is the PC end. Later it changed to the mobile end, and recently it was artificial intelligence. . As we all know, Intel is the forever boss of the PC era. In the 25 years, it basically crushed all competitors, but in the direction of mobile devices, it completely missed it. Therefore, in the same field, it is basically strong and strong, and it is difficult to surpass it. However, being strong in this field does not mean that it will be strong in the next field. It is the easiest to make a dark horse when the "era" is alternated. When judging the value of the company, we must have a judgment on the technical application of the chip. It is very clear where the next so-called vent is.
Tsv Tunnel Spraying Vehicle,Multi-Functional Washing Vehicle,Custom Tunnel Cleaning Car,Automatic Tunnel Spraying Vehicle
CRRC SHANDONG CO., LTD. , https://www.crrcsd.com