Chinas artificial intelligence chip has begun to break out, but there have also been some chaos.

category:Internet
 Chinas artificial intelligence chip has begun to break out, but there have also been some chaos.


Moreover, in the slogan, the words global leading and Chinas first are common. Propaganda is becoming more and more evil, and they all propagate themselves as Chinas first. Under such circumstances, PPTs paper data hanging friends have no meaning at all. What we should really focus on is whether these global leading domestic smart chips are closely integrated with applications. All vendors compete for artificial intelligence chips In recent years, the PC industry has begun to decline year by year, and the smartphone industry has entered the bottleneck with the gradual saturation of the market. In this case, artificial intelligence, Internet of things, cloud computing, big data and other fields are considered to be the next draught. Among them, artificial intelligence is undoubtedly the most beloved by media and capital. It is precisely because of this, foreign factories have launched their own AI chips. Intel launched a crowd of core CPU, which launched a GPU dedicated to artificial intelligence, CEVA and Cadence to launch a DSP for artificial intelligence, and Altra launched a FGPA for artificial intelligence. In addition, Google also launched TPU, IBM launched the true north, the original electric car Tesla also publicized to develop their own AI chips. At the same time, at the same time, the domestic units have launched their own artificial intelligence chips, the CAS calculation has introduced the Cambrian chip, and the middle star micro has also developed the Xingguang intelligent one. At the end of 2017, the horizon launched its own AI vision processor, and Huaxia also released its own Songjiang and Polaris. In addition, some vendors are clearly an artificial intelligent IP that buys friends to integrate into the SoC, but is very artistic in the publicity. It is easy to mislead the public to think that the vendors who buy IP as integration are also developing artificial intelligence chips. Cambrian became Kirin NPU There are more and more boasting about foreign manufacturers, but domestic manufacturers have become more and more boastful. For example, last December, the AI core era product conference held on the horizon asked hundreds of media to help, and announced that the launch of this chip is Chinas first full ball leading high performance, low power, low delay. The embedded artificial intelligence vision processor. In the media coverage, the horizon is also called the worlds leading embedded AI leader. Shortly after the launch of the horizon product, Huaxia also released its own product. Chinas first 64 high-end embedded the Great Wall series CPU/DSP unified processor IP and the Songjiang series embedded artificial intelligence special processor IP, as well as the multi core SoC chip platform based on the fully autonomous IP above - the north star. It is reported that this is the first release of artificial intelligence platform chip with independent intellectual property rights in China. It not only broke the foreign monopoly, but also initially realized the industrialization. It added an important point to the construction and improvement of Chinas autonomous and controllable artificial intelligence industry chain. In fact, there is a certain amount of water in these slogans, such as the so-called Chinas first global leading embedded artificial intelligence processor on the horizon, which is inaccurate. After all, last October, CAS computings embedded artificial intelligence processor, Cambrian 1A, has entered the homes of the common people with HUAWEIs Mate10 listing, and only 2 months later, the horizon released the so-called Chinas first global leading embedded artificial intelligence processor. For example, Huaxia, in essence, is the choice to use the traditional SIMD/DSP architecture to run artificial intelligence. This is the same as last years star intelligence one. Because of the Cambrian, star intelligence, and other products in China, the Cambrian is expected to use the HUAWEI handset to produce tens of millions of levels of production, so called to break the foreign ridge. There is no way to talk about it. Similarly, Baidu launched XPU and Ali just released Ali-NPU. According to media reports, Ali-NPU is still in the design, but publicly claims: the Alibabas Ali-NPU performance will be 10 times the current market mainstream CPU, GPU architecture AI chip, and manufacturing costs and power consumption is only half, and cost performance is more than 40 times. In the case of Ali-NPU being designed, I really dont know how these 10 times and 40 times data are tested. Can it be difficult to get rid of the Dharma courts sweeping monks and Jack Ma TSE? It can be said that due to the heat of artificial intelligence, many manufacturers are eager to win the favor of the capital market. In order to publicize itself in publicity, various labels such as Chinas first, global leading and breaking foreign monopoly have been used up by everyone. The eight immortals cross the sea, each shows the spirit. In these AI chips, there is a rule that the versatility of chips is inversely proportional to the performance and power consumption of AI. Chips like CPU, because of their ten thousand essential oil properties, cause the chip to be very versatile, but in some specific areas, performance and power consumption have no advantage over several other chips. The generality of GPU is worse than that of CPU, but it is not as good as ASIC, so it will be better than CPU in performance and power consumption, but it is obviously lower than ASIC. This is also the reason why Cambrian chip incubated by CAS computing is obviously superior to GPU in performance power ratio. After modifying the traditional DSP processor architecture for digital signal processing, it can also be used to run artificial intelligence, and the existing mature technology can be used. However, it has some limitations in the field of application - it can be used in the convolution neural network (CNN), but it does not help the network of recurrent neural network (RNN) and the long and short term memory network (LSTM) to deal with the network of speech and natural language. In other words, this DSP is mainly used in machine vision field, but it is difficult to be applied to speech recognition, text processing and natural language understanding. Using DSP to do AI chips still has certain commercial potential in the market of artificial intelligence. According to Altras PPT, the FPGA performance power ratio for artificial intelligence can reach 50GFlops/W, so there is a view that FPGA has the advantage of reducing operating costs because of the power saving of the GPUs GPU. However, FPGA also has its own shortcoming, which is the lack of cost performance. Although FPGA has achieved certain achievements in some emerging applications, with the rise of emerging fields, the size and importance of the market has been increasing. Because of the lack of cost-effective reasons, FPGA will eventually be replaced by a dedicated chip. However, the higher the performance is, the lower the problem is, the real important is the application and industrial ecology. The reason why GPU is becoming mainstream is that the ecosystem is well developed and the development is convenient. Artificial intelligence should not be overhyped Because artificial intelligence is very hot, many companies start to play the concept, whether it is a bitcoin or DSP, and become an artificial intelligence chip company. Up to now, more than 40 AI chips have been financed or are being financed by China, and most of them were established after 2015. At present, artificial intelligence has been overhyped, and some media have advocated that China rely on artificial intelligence to overtake. But in fact, artificial intelligence chips are only accelerators to solve specific problems, and can not replace CPU, GPU, FPGA, DSP, NANDFlash, DRAM and other types of chips. Some media claim that China should vigorously develop AI chips and rely on artificial intelligence to break the monopoly of X86 and ARM. This argument is obviously lacking in basic industry knowledge. The effect of AI chips on the industrial chain is limited. For example, in order to realize the pull effect on the upstream and downstream of the industrial chain, the domestic chip design company, at the cost of sacrificing the chips performance, chooses the chip manufacturers production chip within its territory; the domestic wafer plant can make certain sacrifices and choose the products of the domestic equipment dealers and the raw materials manufacturers... This can form a positive cycle. If the chip design companies in China always look for TSMC, the domestic Fabs prefer the equipment of ASML, applied materials and other international factories. Well, Chinas chips are hard to resist. However, such as Ali, Baidu, the horizon, and the bit continent are very commercial companies, must be the commercial interests of priority, the inevitable choice of TMV more mature technology. Even if the order is handed over to SMIC, it is also because SMIC has greatly improved in terms of yield and so on. In a word, only the icing on the cake will not be delivered in the snow, and the most difficult thing in the middle core international is that the new technology can not find the customers willing to go down in the early stage, and can not use the capacity to spread the cost. This will fall into a dead cycle. At present, one of the main drivers of breaking the cycle is not domestic enterprises, but the US Qualcomm Corp. It is not that Qualcomm harbors a red heart, but it is only monopolized by the NDRC that the order is given to SMIC. In addition, not only Ali, Baidu and other Internet Co will not help, but HUAWEI, ZTE, millet and other companies also give priority to TSMC. Therefore, the large amount of capital invested in artificial intelligence is likely to become a carnival of capital draught. It can not only break the monopoly of Western companies on chips, but also does not necessarily play a role in promoting the industrial chain. And the real short board of Chinese chip, such as CPU, GPU, FPGA, RF, AD/DA, and all kinds of development tools, may be faced with the situation of nobody. No help for repeating the mistakes of the ZTE incident. The source of this article: editor of science and technology force: Hou Wei Cheng _NT4124 At present, one of the main drivers of breaking the cycle is not domestic enterprises, but the US Qualcomm Corp. It is not that Qualcomm harbors a red heart, but it is only monopolized by the NDRC that the order is given to SMIC. In addition, not only Ali, Baidu and other Internet Co will not help, but HUAWEI, ZTE, millet and other companies also give priority to TSMC. Therefore, the large amount of capital invested in artificial intelligence is likely to become a carnival of capital draught. It can not only break the monopoly of Western companies on chips, but also does not necessarily play a role in promoting the industrial chain. And the real short board of Chinese chip, such as CPU, GPU, FPGA, RF, AD/DA, and all kinds of development tools, may be faced with the situation of nobody. No help for repeating the mistakes of the ZTE incident.