China's Efforts to Build the Semiconductors at AI's Core
As U.S.–China tensions escalate, China's project to develop independent tech takes shape
Flickr / Fritzchens Fritz
Dec. 7, 2018
The Chinese leadership in recent months renewed its emphasis on artificial intelligence (AI) technologies as essential to the country’s development goals. Among the most crucial technologies fueling AI systems are semiconductors, an area where Chinese companies, including those such as Huawei that are frequently subject to U.S. government scrutiny, are hard at work.
It’s not just that Chinese companies would like to compete in these technologies. The Chinese government wants independence from international suppliers. Last month, in the first known Politburo study session on AI, General Secretary Xi Jinping said China needed to “ensure that critical and core AI technologies are firmly grasped in our own hands.”
And in a wide-ranging speech to the National People’s Congress Standing Committee, a leading scholar from the Chinese Academy of Sciences (CAS) assessed China’s strengths and weaknesses in AI fields, calling for China to “construct an independent and controllable innovation system.”
The scholar, Tan Tieniu, made explicit what has long been apparent: U.S. government actions toward the Chinese network infrastructure and smartphone giant ZTE have increased Chinese resolve. ZTE was threatened with destruction and actually shut down operations after U.S. authorities moved to ban it from purchasing crucial components after it said ZTE breached a settlement in a previous case in which ZTE admitted to violating U.S. sanctions on Iran and North Korea.
Though many in China thought ZTE had been reckless, the fact that the U.S. Commerce Department had the leverage to practically destroy a major Chinese company focused many minds. Only after Xi intervened with Trump was the company saved.
In this context in April, Xi reminded audiences of the value of “self-reliance” and touted “indigenous innovation” in “core technologies” at the important National Cybersecurity and Informatization Work Conference.
“The U.S. ban on ZTE,” Tan said in his speech, “fully demonstrates the importance of independent, controllable ‘core-, high-, and foundational’ technologies.” Although the “core-, high-, and foundational (核高基)” formulation dates back at least to 2006, Tan coined a “new core-, high-, and foundational” concept and described it as follows:
“‘New’ refers to the new type of open innovation ecosystems such as civil-military integration, and the integration of industry, academia, and research. ‘Core’ refers to core and critical technologies and components, such as advanced machine learning technology, robust pattern recognition technology, and low-energy-use intelligent computing chips. ‘High’ refers to high-end integrated application systems and platforms, such as machine learning software and hardware platforms, and large-scale data platforms. ‘Foundational’ refers to theories and methods with great innovative significance and potential for driving technological progress, such as brain-computer interfaces, brain-like intelligence, etc.”
Semiconductors: The technologies at the core of core technologies
Whether it’s “core-, high-, and foundational” or just “core,” semiconductors are an essential part of Chinese efforts to develop indigenous or independent technologies—especially in AI fields.
The influential New Generation Artificial Intelligence Development Plan (AIDP), issued in July 2017 and translated by DigiChina, noted China was “lacking major original results in … high-end chips.” The AIDP called for breakthroughs in “intelligent computing chips and systems,” “high energy efficiency, reconfigurable, brain-inspired computing chips” and “new-model perception chips,” as well as those adapted for Internet of Things devices.
Half a year after the AIDP, Chinese efforts in semiconductors were still nascent and uncertain (see section 5, “Semiconductors: Attempting to Chart a Different Course” here).
Within the last six months, the major AI platform companies Baidu and Alibaba have both announced plans to develop their own AI-optimized semiconductors (or ICs, for “integrated circuits”). The companies are likely to follow leading Western firms, with “fabless” IC design shops—operations that turn out chip designs that are then fabricated by leading IC companies such as Taiwan Semiconductor Manufacturing Corporation (TSMC). All of China’s leading AI chip players are signed up with TSMC for production--Taiwan’s semiconductor prowess is China’s secret weapon in the coming semiconductor and AI chip wars.
Chinese Semiconductor Efforts So Far
|Firm||Focus on IC design||Listed partnerships||Notes|
|LARGE TECH FIRMS|
|Huawei Technologies||Huawei's IC design arm HiSilicon has already designed and had fabricated the Kirin 9XX series of AI neural network processors. It also designed the Ascend 910 processor for data center AI applications, and the Ascend 310 for edge servers.||HiSilicon-designed AI processors are primarily fabricated by the Taiwan Semiconductor Manufacturing Corporation (TSMC).||HiSilicon designs exclusively for use in Huawei products.|
|Baidu||Announced in July 2018, Kunlun chips for edge and cloud computing are designed for autonomous vehicles and data centers. These chips are not yet available in significant commercial quantities.||Baidu has partnered with 1) Intel to optimize Xeon processor running PaddlePaddle, the firm's deep learning framework; 2) Nvidia to customize PaddlePaddle for Volta GPUs.||The Kunlun, at least in theory, is designed to compete with Google's Tensorflow Processing Units (TPUs).|
|Alibaba||Formally established a semiconductor business operation to produce AI chips. AliNPU is designed for autonomous vehicles, smart cities, and smart logistics, and will be overseen by the Damo Academy. Subsidiary Pintouge will focused on customized AI chips and embedded processors.||Alibaba has so far not partnered with foreign hardware firms. It has a Machine Learning Platform for AI that will likely serve as the software component of its AliNPU or other processors designed for data centers.||In April, Alibaba acquired Hangzhou C-Sky Microsystems, an IC design company.|
|Cambricon Technologies||Cambricon builds core processor chips for intelligent cloud servers, intelligent terminals, and intelligent robots. Its MLU100 chip was released in May for cloud computing along with the IM chip for edge computing.||Cambricon's chips are all fabricated at TSMC. The Lenovo Thinksystem SR650 server is based on the MLU100, as is Sugon's PHANERON series. iFlytek is also collaborating with Cambricon.||Cambricon has two product lines: terminal ICs and server ICs. The Cambricon-1A processor, introduced in 2016, is the world's first commercial deep learning processor for smart phones, wearable devices, unmanned aerial vehicles, smart driving, and other types of terminal equipment.|
|Horizon Robotics||Founded in 2017 by the former head of Baidu's AI unit Kai Yu, Horizon focuses on imaging and facial recognition. Its Sunrise processor runs facial recognition algorithms and is used in smart cameras.||Most of Horizon Robotics chips have been fabricated at SMIC. Its Journey 1.0 application-specific IC (ASIC) for advanced driver assistance systems (ADAS) will use TSMC.||Other founders include Huang Chuang from Baidu's AI business unit, and Yang Ming from Facebook's AI research team. Chip development is headed by Zhou Feng, former principle chip designer for HiSilicon.|
|DeePhi||The firm, now part of Xilink, focuses on field programmable gate arrays (FPGAs) for AI applications. Its DeePhi Deep Learning Processing Unit (DPU) is the hardware basis for development architectures.||DeePhi’s software engines for optiimizing algorithms were deployed on Xilinx FPGAs, as one of the company’s earliest product solutions.||DeePhi features the Aristotle Architecture for convolutional neural networks (CNNs), and the Descartes Architecture for recurrent neural networks (RNNs). The DNNDK™ (Deep Neural Network Development Kit) is designed as an integrated framework to accelerate deep learning (DL) application development and deployment on the DPU platform.|
|Yuntian Lifei (Intellivision)||The firm develops visual recognition and big data analysis ICs and systems. Its NNP100 and NNP200 processors were developed in-house. These are edge-based AI chips for use in smart vision and security monitoring systems.||Intellifusion also has partnered with Huawei on an AI Vision Platform. The NNP200 (22 nm process) is likely fabbed by TSMC.||The firm is very new, having been founded in Shenzhen in 2014 by doctoral degree returnees from the United States. The company claims to be responsible for both the processor technology, supporting software, and big data analytics.|
|Enflame (Suiyuan) Technology||A very new company, founded in March 2018, Enflame has Beijing- and Shanghai-based R&D centers. It is focusing on cloud-based deep learning chips that appear to be designed to compete with Nvidia GPUs.||In October, Enflame purchased multiple licenses from Arteris, specifically the FlexNoC interconnect IP for use as the on-chip communications backbone of its AI training chips for cloud datacenters.||Suiyuan Technology CEO Zhao Lidong(赵立东), formerly worked for AMD China. Zhao says the company will follow the Chinese government’s Development Plan on the New Generation of Artificial Intelligence, and aims to become the AI chip solution and technology leader in the Chinese market.|
|Iluvatar CoreX||Founded in 2015 by semiconductor engineers from Silicon Valley and China, the firm is focusing on high-end cloud computing chips (GPUs) and computing infrastructure software called SkyDiscovery.||Illuvatar CoreX has also licensed Arteris IP FlexNoC Interconnect technology for deep learning SoC applications.||The companies founders claim to have been working together for more than 20 years at companies including Oracle, AMD, and ATI. The firm has R&D centers in Nanjing, Shanghai, Beijing, and Silicon Valley.|
|Bitmain||Bitmain last year launched the BM1680 AI chip. The firm is billing the BM1680 as Tensor Processing Units (TPUs). The BM1682 released this year is a dedicated processor for image and video processing.||Bitmain's chips are fabricated at TSMC. In September it announced that the cryptomining chip BM1391 will be based on the TSMC 7 nm process.||Bitmain began designing ASICs for bitcoin mining and currently dominates that market. The firm turned to designing AI-optimized chips in 2017.|
|Canaan Creative||Canaan is located in Beijing in "Silicon Valley Bright City." It focused on blockchain servers and solutions for repetition ASIC chips.||Canaan Creative's latest ASICs will be manufactured based on TSMC's 7 nm process.||The firm was acquired in June 2018 by Shandong Luyitong Intelligent Electric.|
It will take some time to determine whether Baidu’s or Alibaba’s chip design efforts can really compete with established leaders such as Nvidia or Intel, particularly for data center chips running AI applications in the cloud.
DigiChina wrote in February, developing, scaling, and successfully iterating complex hardware designed to run AI algorithms will be very difficult. Western firms such as Nvidia have a huge head start, and it remains unclear if Chinese firms, including large and well-funded players such as Alibaba and Baidu can attract and keep sufficient numbers of design engineers to break into a significant market share for areas such as GPUs.
The only company with sufficient design and engineering talent to compete with the large international firms is Huawei, due to its subsidiary HiSilicon. Huawei, of course, has long been under scrutiny by the U.S. government as a national security threat and sanctions violator—and the arrest of its CFO over U.S. sanctions is a major new development. The Chinese government’s efforts to ensure autonomy in core technologies, meanwhile, are aligned with Huawei’s success, making the company’s development efforts and products especially sensitive in the coming months.
Acknowledgment: Thanks to Kevin Allison (@KevinAllison) for extremely valuable research support.
Disclosure: China Digital Economy Fellow Paul Triolo also works at Eurasia Group, which provides analysis and serves clients in fields related to DigiChina’s work. DigiChina contributors commit that their work for New America is not designed to serve the interests of outside clients, named or unnamed.