Global AI chip list: HUAWEI twelfth, seven Chinese companies finalists Top24

category:Internet
 Global AI chip list: HUAWEI twelfth, seven Chinese companies finalists Top24


NetEase news May 3rd, recently, market research company CompassIntelligence published the latest research report, the top 15 top AI chip enterprise ranking in the world, the top three are Ying Weida (Nvidia), Intel (Intel) and NJI Pu (NXP), HUAWEI is ranked twelfth, become the only vaccine of TOP15 in China. It is understood that the list of A in the AI chipset index of this report includes companies that provide AI Chipset Software and hardware components. AI chipset products include central processor (CPU), graphics processor (GPU), neural network processor (NNP), specific integrated circuit (ASIC), field programmable gate array (FPGA), simplified instruction set computer (RISC) processor, accelerator and so on. Some chipsets deal with edge processing or devices, some servers for cloud computing, and others for machine vision and automated vehicle platforms. Some of these products are AIs computing framework, while others are AI training platforms. The report also mentioned that over the past three years, in addition to its own research and development investment, a total of $60 billion has been invested in the field of artificial intelligence. We have seen more than 1700 start-ups interested in AI chips, and, of course, the industrys demand for AI chips is also increasing. At the same time, in addition to the TOP15 rankings, CompassIntelligence also assessed as many as more than 100 chip companies, which included 24 companies in the final ranking, including the envier, Intel, IBM, Google, apple, AMD, ARM, Qualcomm, Samsung, NJI Pu and so on. It is worth noting that the Chinese enterprise HUAWEI Still the twelfth place, the Cambrian and the horizon are twenty-second and twenty-fourth respectively. In the list of Top24, a total of seven Chinese companies were selected. MediaTek ranked fourteenth. Imagination ranked fifteenth; Verisilcon ranked twenty-first; The horizon (Horizon) ranked twenty-fourth. In October 2004, HUAWEI founded the Hass company, whose predecessor was the HUAWEI IC Design Center, which also officially opened the way of R & D of HUAWEIs mobile phone chips. In 2009, HUAWEI launched its first open market - oriented K3 processor, which was positioned to compete with the exhibition and the coalition to compete in the Shanzhai market, and HUAWEIs own mobile phone was not used. Because the K3 product is not mature enough and the sales strategy is not suitable, the chip has not been successful. This is also the first smart phone processor in China. In 2012, China launched K3V2 processor for Hass. This time it was used in its own mobile phones, and it was the flagship Mate1, P6 and other models. In 2012, the mobile phone processor has opened the multi-core process, and K3V2 has become the second quad core processor in the world. Then, Kirin 910 is the first SoC of the Hester. If CPU is the brain of the mobile phone, the SoC is a system that integrates various functions and assigns tasks to them. A mobile SoC includes the baseband (Baseband), the graphic processor (GPU), the digital signal processor (DSP), the image signal processor (ISP) and so on. For modules. Since K3V2, some of the HUAWEI phones, especially the flagship machines, have been using their own Hass chips, and more importantly, the HUAWEI flagship bindings have to make rapid progress and steady supply. We can see that in 2014, HUAWEIs R & D investment was more than the sum of 400 A shares. In 2017, HUAWEIs R & D cost reached 89 billion 700 million yuan, much higher than apple and Qualcomm. Over the past ten years, HUAWEI has invested more than 394 billion yuan in research and development. Thats not surprising. After more than ten years of development, in September 2017, HUAWEI officially launched its latest AI chip kylin 970 (Kirin970) at the Berlin international electronic consumer goods exhibition (IFA), Germany. Kirin 970 uses the industrys high standard TSMC10nm process, which integrates 5 billion 500 million transistors on a nail size chip, reduces power consumption by 20%, and achieves a 1.2Gbps peak download rate. Kirin 970 integrates NPU dedicated hardware processing unit (Cambrian IP), innovating the HiAI mobile computing architecture, and its AI performance density is much better than CPU and GPU. Compared to the four Cortex-A73 cores, the new heterogeneous computing architecture has about 50 times energy efficiency and 25 times performance advantages in handling the same AI task. Moreover, HUAWEIs second generation AI chip Hirst Unicorn 980 will also be formally produced in the current quarter, using TSMC 7Nm process technology. This processor will be equipped with the second generation of NPU, which will support more scenario applications on the basis of previous generations, and the performance of NPU will be increased by more than 2 times. Core movement in China In May 3rd, Cambrian issued a new generation of terminal IP products in Shanghai, using the terminal chip Cambricon1M of 7Nm technology, the first cloud smart chip MLU100 and the cloud end intelligent processing card carrying MLU100. The third generation machine learning terminal processor 1M has 10 times higher performance than the previous Cambrian 1A. In terms of configuration, Cambrian 1M is produced by TSMC 7Nm process, and its 8 bit operation efficiency is 5Tops/watt (5 trillion times per watt). Cambrian provides three processor cores of 2Tops, 4Tops and 8Tops to meet the needs of different levels of intelligent processing in different scenarios. MLU100 uses the latest Cambrian MLUv01 architecture and the TCP 16nm process, which can work in two different modes of equilibrium mode (primary frequency 1Ghz) and high performance mode (main frequency 1.3GHz). The equivalent theoretical peak velocity can reach 128 trillion fixed-point operations and 166 trillion and 400 billion fixed-point operations respectively, and their power consumption is 80W and 110W. The Cambrian introduction, MLU100 cloud chip also has high throughput, can support all kinds of deep learning and common machine learning algorithms. They also put forward the idea of end cloud collaboration, that is, MLU100 cloud chip can adapt to the Cambrian 1A/1H/1M terminal processor and cooperate with the complex intelligent processing task. . At the end of last year, the horizon launched embedded artificial intelligence chips - the Journey 1 processor and the Sunrise 1 processor for the smart camera. Yu Kai thinks: the horizon sees the future of the artificial intelligence processor, in fact it is also the commanding point of our countrys scientific and technological competitive power. If the future of Chinas artificial intelligence industry is to take off, to take off, it must have an aero engine. What is this aero engine? It must be an artificial intelligence processor. There are still a lot of start-ups joining the war. Cloud sound and Rokid have announced the completion of the chip research and development. Cloud knows the upcoming release of the AI chip, which is based on the Unisound AI instruction set and DSP instruction set, combined with the voice application scene, with the microphone array signal processing, speech recognition and speech synthesis as a whole. Chip architecture. According to the introduction, this AI chip ensures the efficiency by the programmable interconnected matrix between the operation units, and uses a multistage multi group multi port Memory architecture to ensure the enhancement of the data bandwidth in the chip and reduce the power consumption of the chip. In the aspect of architecture flexibility, the main control CPU is connected with the AI accelerator internal RAM through Scratch-Pad, providing an efficient data channel between the CPU and the AI accelerator, so that CPU handles the results of the AI accelerator operation two times. In addition, the programmable interconnected matrix architecture that connects each operation unit provides the function of the extended operation instruction to further enhance the flexibility and extensibility of the hardware architecture. The rest of the exploration of the chip architecture, including multilevel multi mode wake-up, from energy detection to human voice detection to wake-up words detection, and customized PowerDomain for voice devices and using scene, minimizes chip power consumption. Domestic media analyzed four big mountains facing domestic chip makers: 1, the accumulation and endurance of long-term R & D investment. It is reflected in the design of microarchitecture, the lack of design ability of the underlying operating system, and the general CPU without its own micro architecture (most of the domestic PC/ server operating systems are still based on Linux. In these respects, foreign ARM and other manufacturers are actually going through more than 20 years of research and Development), or the rapid introduction and snatching of the top. Sharp chip design talent. 2, realize the positive cycle of heavy capital input and high output. 3, short term performance and stability surpass foreign competitors. 4, the cultivation of the ecology of the hardware developers. Intel and MS have rarely adopted such cross level strategic operation in domestic universities for many years to develop curriculum system, certification system and ecological cultivation system. Pay attention to NetEase intelligent public address (smartman163), to interpret the big events, new ideas and new applications of large companies in AI field. Source: NetEase intelligent responsible editor: Wang Chao _NT4133 1, the accumulation and endurance of long-term R & D investment. It is reflected in the design of microarchitecture, the lack of design ability of the underlying operating system, and the general CPU without its own micro architecture (most of the domestic PC/ server operating systems are still based on Linux. In these respects, foreign ARM and other manufacturers are actually going through more than 20 years of research and Development), or the rapid introduction and snatching of the top. Sharp chip design talent. 2, realize the positive cycle of heavy capital input and high output. 4, the cultivation of the ecology of the hardware developers. Intel and MS have rarely adopted such cross level strategic operation in domestic universities for many years to develop curriculum system, certification system and ecological cultivation system. Pay attention to NetEase intelligent public address (smartman163), to interpret the big events, new ideas and new applications of large companies in AI field.