Reporter | Cai Haoshuang Li Zhonghao Editor | Zhao Li
Alibaba's acquisition of Zhongtian Micro, independent research and development of 'China Core', opened the news of the 'first shot' of the Sino-US trade counterattack war, and injected a strong 'core' needle into China for 'core pain'. Public opinion seems to see it again. The opportunity for China's chip to overtake the curve. In fact, we should face it squarely:
The Chinese chip industry still has a long way to go. Although it has been working hard, the high barriers in the chip industry make it difficult to bend overtaking in the short term.
The reason why this is said is because the integrated circuit industry chain is extremely complex, including software (EDA tools), design, manufacturing, packaging and testing, materials, and equipment. Autonomous design does not mean independent manufacturing. Having hardware does not mean that software can also be produced. Low-end chips do not equal high-end chips. The market potential of dedicated chips does not mean that they can replace general-purpose chips.
The problem is that in the days when the 'chip' is being swiped, what we most need to know may be the 'chip' we are talking about.
High school low-end chip classification
Digital chips, analog chips are divided into high, medium and low chips. In simple terms, high-end chips solve many complex problems, mid-end chips solve a single complex problem, and low-end chips solve a simple problem.
Among the more familiar chips, CPUs, GPUs, deep learning chips based on FPGA architecture (Field Programmable Gate Arrays), and ASICs (Application-Specific Integrated Chips) are all high-end general-purpose chips. A slight difference is that they are based on FPGAs or ASICs. The deep learning chip is mostly used in the field of artificial intelligence, especially for smart cameras, and therefore more like a programmable dedicated chip.
Alibaba Dharma announced yesterday that it is developing a neural network chip - Ali-NPU, which is a dedicated chip for imagery video analysis, machine learning and other AI reasoning calculations. Since 2017, with artificial intelligence gradually moving toward the application layer, The AI chip has become a wave of entrepreneurship. The Cambrian, Horizon, Shenzhen Science and Technology, which emerged in this wave, are not based on FPGA or ASIC-specific chips. It is worth mentioning that the current fastest The chip is applied to the Cambrian period on Huawei mobile phones. It is a 'national team' with the background of the Chinese Academy of Sciences. FPGAs and ASICs are still all dependent on imports.
Compared to general-purpose chips, the threshold of AI chips is relatively low. 'The overall gap in the chip field in China and the United States is more than 10 years. The AI related field is a rare industry with the same starting line.' Shen Jian Technology CEO Yao Song told the founder of China.
Song Hui Innovation Fund Investment Director Liang Hui once worked in Huawei's chip business. According to Liang Hui, to determine whether these AI chips are subject to foreign parties, it depends on whether they use the CPU's IP core in the algorithm. 'As far as I understand, there are The AI chip scheme includes the core of ARM in the algorithm, and the IP of such core is certainly still used in the foreign IP. Some pure ASIC algorithms are related to the nucleus. In this case, if the entire design is mastered In your own hands, you will not be affected by foreign countries.
According to the search for China Maker Reporter, at present, the CPU IP Core technology is mainly provided by British and American companies such as ARM, AMD, IBM, etc. The previously mentioned Alibaba acquired Zhongtianwei is currently the only large-scale independent production in mainland China. Embedded CPU IP core company. Although this move demonstrates the determination of Jack Ma's self-developed chip, but a chip industry veteran believes that Zhongtian Micro's CPU IP core and ARM, IBM is obviously not in the same grade. 'As far as I know, the domestic present CUP wants to benchmark against Intel, it is still difficult to match.
Divided from the application scene, mobile phone mobile processor, advanced automatic driving master chip, chips for the development of 5G communications, etc., are all high-end chips.
In terms of mobile phone chips, MediaTek and Qualcomm have almost monopolized the mobile phone chip market. Almost all of the Android system's flagship phones use Qualcomm's Opteron processors, while mid- to low-end Android models generally use MediaTek processors. Apple, Samsung, and Huawei Special, use its own processor. There are close to Huawei chip insiders told reporters seeking Chinese founders, the first generation of Huawei Kirin chips is not yet mature, but Huawei's overall design of mobile phones cover up the shortcomings of the chip, this allows the chip can be used to land , gave the opportunity to iteratively upgrade the chip products. And this process, Huawei spent nearly 20 years.
In the area of autonomous driving, Tumon’s future CEO Chen Mo told reporters looking for China’s makers. Only the Nvidia’s GPU can be used for the master chip of automatic driving above the L4 level. Tucson’s future solution is calculated by the master control calculation unit. There is no increase in end-side chips. 'It is difficult for domestic companies to make master chips that meet the L4 level of autopilot, but if we do ADAS-specific chips, I think it is still possible. There are few general-purpose chips in China, and they generally do AI-specific chips. . '
Universal chip VS special chip
The dispute about universal chips and dedicated chips has never stopped.
There have been views that dedicated chips are the new growth point in the future.
Those who hold this view believe that although universal chips have great advantages in terms of versatility and modificability, in terms of pertinence and power consumption, general-purpose chips need to incorporate many reserved modules at the beginning of design. Therefore, the power consumption will be very high, and it will be expensive to sell. The dedicated chip is designed for a single function. Many things have been hardened and blocked, so it is fast, low power consumption, low cost, and low power bills. .
But Chen defaulted to the fact that dedicated chips have a long development cycle and their use is not flexible. 'GPU can be added or changed by pure software, upgrading functions. Special chips can't. 'So, dedicated chips are only suitable for stable single tasks. Although dedicated chips The market is vast, but there are still challenges for Chinese companies.
Because even mastering chip design, packaging and measurement technologies, but in the software (EDA tools), manufacturing, equipment, etc., high-end 'China's core' is still facing many difficulties.
'EDA software tools are still in the hands of foreign companies such as CADENCE, SYNOPSYS. As far as I know, SYNOPSYS has stopped selling to ZTE.'
In terms of equipment, the lithography machine is one of the core devices for chip manufacturing. According to media reports, ASML, the world's largest supplier of chip lithography equipment, is one of the few manufacturers. But this is the core device, but it is for China. Forbidden to sell. Even if it were sold to China, it would be a few generations behind.
Restricted by the manufacturing industry, even Kirin 970, which is comparable to Qualcomm's Opteron chips in design, has to be built by TSMC in the manufacturing process. It is understood that the Cambrian, Horizon, etc. have also chosen foreign manufacturers. Cooperation.
The high-end chips needed by ZTE are the areas that can be covered by long-term huge investment through decades of technology accumulation and R&D.
U.S. NVIDIA and Intel as GPUs and CPUs are the top players in the two most typical high-end chips, and both have achieved long-term technological accumulation. They are the key raw materials in the microelectronics industry: 'oil' and 'salt'. People.
The explosion of the AI concept for the past two years fired a large number of 'artificial intelligence dedicated chips' such as the NPU and TPU. The industry initially used GPUs as an artificial intelligence training and computing platform, but due to excessive power consumption, it was wasted. The so-called 'artificial intelligence dedicated chip' was produced. That is to say, these chips actually only solve the problem of induction of machine learning algorithms. Therefore, the US's ban on ZTE's ban has only become so impactful.
Ali today announced the acquisition of wholly-owned Zhongtian Microsystems Co., Ltd., the only company in mainland China with its own embedded CPU IP Core, and the technology related to 'IP Core' is the core technology for independently researching and developing high-end chips. , Ali's competitors are NVIDIA, Intel, AMD and other first-tier manufacturers.
In any case, this move also shows that Ali officially entered this field and supports domestic high-end chip manufacturing.