[ Security Exhibition Network Enterprise Attention ] For practitioners of the information technology industry, 2019 means a turning point. After Huawei cut the confession, there have been waves of independent core creation. For a while, chips have become the focus of everyone's attention. In fact, with the development of China's semiconductor industry, a number of outstanding local chip companies have moved from behind the scenes to the front. Among them are old players represented by Huawei Hisilicon and Ziguang Zhanrui, as well as emerging players such as the Cambrian and Jianan power.
In the rising Chinese AI chip industry, "curve overtaking" will inevitably roll over, and only by mastering the hard core technology can we get ahead. This article will review ten representative "Chinese chips" from the four camps of mobile phone chips, edge computing, machine vision and machine hearing.
5G mobile camp
Huawei Hisilicon: Kirin 990 5G
In the hot discussion triggered by the confiscation of supply this year, Huawei Hisilicon, which had been in storage for many years, turned straight overnight, allowing people to see a technical hole card in the game of great powers. At the September International Consumer Electronics Show in Berlin, Huawei Hisilicon mobile chip Kirin 990 5G was officially released. This chip is the industry's first full Netcom 5G SoC, with a 5G NR downstream peak rate of 2.3Gbps and an upstream peak rate of 1.25Gbps. It is the only SoC that supports 5G dual-mode so far. This also means that on the global 5G chip track, Kirin 990 5G will take the lead to get the first-mover advantage.
Purple light exhibition sharp: Tiger 710
In the field of 5G mobile chips, the pace of Ziguang's sharpening followed Huawei. Recently, 5G prototypes equipped with Zhanruihu 贲 710 + Chunteng510 have passed the comprehensive verification of the China Telecom Research Institute's Tyre Lab. The prototype has performed well under the test of a total of 80 testing items such as performance and functions, information security, interconnection and electromagnetic compatibility performance, which means that the Ziguang Zhanrui 5G chip has reached the commercial level.
Second, the edge AI camp
Ali flat-headed brother: Xuantie 910
Since the merger of Zhongtian Wei and Dharma Institute to form a flat-headed brother, it has been less than one year. On July 25th, Ali Pinto released the RISC-V architecture chip Xuantie 910. It is said that the chip has two major innovations: the first is the industry's first RISC-V processor that implements 2 memory accesses per cycle; the second is based on RISC-V, which expands more than 50 instructions and systematically enhances the RISC-V Computing, storage, and multi-core capabilities. In terms of performance, the Xuantie 910 is 40% faster than the mainstream RISC-V instruction in the world. It is a chip with strong computing power in the current RISC-V camp. However, it is unclear whether the Xuantie 910 has already been taped out.
Jia Nan: Kanzhi K210
In the RISC-V alliance camp, the strength of another company cannot be underestimated. As an important member, Jia Nan is one of the top five manufacturers using RISC-V chips in the alliance. Canaan officially launched the research and development of AI chips in 2016, and released the first generation of AI chip Kanzhi K210 two years later, which is also the industry's first independent intellectual property edge AI chip based on RISC-V architecture.
The chip can run at 0.3W low power consumption, and provides 1TOPS computing power, which has high energy efficiency. In addition, K210 has three characteristics of integrated audio-visual design, strong programmability and independent IP. First, the K210 has multi-modal recognition capabilities for machine vision and machine hearing. Second, the chip is equipped with a FPIOA field-programmable IO array. One chip can burn different algorithms and has strong programmability. Finally, the neural network accelerator KPU of the chip is completely based on the independent research and development of Canaan, which can efficiently complete tasks such as image classification, face detection and recognition in edge-side scenes. As of September 30 this year, the chip has shipped more than 53,000 units, the Q3 quarter increased more than doubled.
It is reported that the Kanzhi K510, the second-generation chip of the Kanzhi series, is also expected to be released and mass-produced in Q1 next year. The second-generation chip still uses the RISC-V dual-core processor architecture. In order to make up for the lack of KPU's algorithm versatility, the K510 will incorporate a self-developed computing module GNNE in its design. On the other hand, the second-generation chip will support 5G communication modules and be equipped with a 5-channel depth camera
interface, and will be applied in the fields of smart retail, ADAS assisted driving and finance.
Cambrian: Siyuan 220
Cambrian was born in the Institute of Computing Technology of the Chinese Academy of Sciences, and is one of the pioneers of the domestic processor architecture. In November this year, the Cambrian released the chip Siyuan 220, which means that it has realized a comprehensive layout from mobile terminals, the cloud to the edge. As the first chip of the Cambrian in the edge area, the Siyuan 220 is equipped with the Cambrian latest generation of intelligent processor MLUv02, with a computing power of 32TOPS, which is comparable to NVIDIA Jetson Xavier NX in performance.
In terms of software, Siyuan 220 continues to support the Cambrian Neuware software tool chain, supporting mainstream programming frameworks in the industry, including Tensorflow, Caffe, mxnet, and pytorch. Based on Siyuan 220, a small-sized M.2 accelerator card was launched to the market in the early Cambrian period, and a product form with higher computing power will also be launched in the future, which is also worth looking forward to.
Third, the speech processing camp
HKUST News: CSK400X Series
Voice is becoming one of the mainstream human-computer interaction methods in the future. In this field, HKUST Xunfei is undoubtedly the king of the domestic market. The biggest highlight of this year's HKUST Xunfei is the CSK400X series of dedicated voice chips for the home appliance industry. Smart home
is a hot spot in the current AI consumer market. The release of the CSK400X series can be said to be targeted. This series integrates HKUST Xunfei's best speech algorithm, and the entire chip computing power reaches 128 GOPS. Like Xiaomi, InfoTech also focuses on its core strategy in the AIoT field. Its open platform has provided 267 AI capabilities and solutions, supporting speech recognition in up to 59 languages and 28 dialects.
Yun Zhisheng: Swift
Of course, more than HKUST Xunfei wants to take a share in the trillion-level AIoT market. Yunzhisheng also uses the voice chip as a cut to enter the market strategy of AIoT. Its first-generation chip "Yuyan" is targeted at two human-computer interaction scenarios of smart homes and smart speakers. Adopt CPU + uDSP + DeepNet architecture, support 8/16 bit vector, short-distance operation, based on deep learning network architecture. It is reported that the second- and third-generation chips of Yunzhisheng will be extended to the fields of smart cities
and in-vehicle computing, etc., which can be regarded as efforts to break through the voice circle and expand from depth to horizontal.
The core of speech algorithm manufacturer Spitzer started in March 2018. Earlier this year, Spitzer officially released the first AI chip TAIHANG, firing the first shot into the voice AI chip. Algorithm manufacturers making cores will bring a different perspective to the chip industry, which is also reflected in the design of its chip TAIHANG. The chip's main algorithm is deeply integrated with the chip architecture. It is understood that TAIHANG has a complete AI algorithm instruction set and hardware acceleration engine, dual-core enhanced DSP, and a customized instruction set with voice signals, without the need for off-chip storage.
Fourth, the visual processing camp
Skylight Flying: DeepEye1000
Visual processing is another direction in the current AI field. Yuntian Lifei DeepEye1000 is specifically designed for end-to-side vision applications. The chip has a built-in neural network processor with independent intellectual property rights, supports flexible and programmable computing streams, and is equipped with a deep custom instruction set independently developed by Yuntian Lifei, which can achieve high-performance, low-power consumption CNN network model inference calculation acceleration.
Yitu Technology: QuestCore Chip
In the past 4 years, the accuracy of face recognition algorithms based on images has been improved by 100,000 times. Against this background, AI is pushing the computing industry into an "algorithm as a chip" era. According to the figure, questcore uses a self-developed chip architecture. In addition to the acceleration of machine vision inference operations, one of its biggest features is flexibility and scalability. In the design, it fully considers different application scenarios and can take into account cloud and edge visual inference computing requirements.
2019 will soon be over. The turbulence of the technology industry in the past year has made people deeply aware of the importance of core technology being autonomous and controllable. At the crossroads of the new decade, as AI technology enters the large-scale landing stage, finding the optimal computing architecture has become the goal of all manufacturers. A number of local core manufacturers such as Huawei Hisilicon, Cambrian, and Jianan have already been on the road. In the future journey, as local chip manufacturers break ground, domestic market demand will also enter the domestic production period, and the next miracle may be born between them.