At the full-link conference, Huawei divided the company's artificial intelligence development strategy into five parts:
1. Fundamental research on investment: constructing data efficient (less data demand), energy efficient (lower computing power and energy consumption) in the fields of computing vision, natural language processing, decision-making reasoning, safe and credible, automatic autonomy Machine learning basic skills.
2, to create a full-stack solution: to create a full-stack, independent and collaborative, full-stack solution for cloud, edge and end, providing ample, economical computing resources, easy to use, efficient, full process AI platform.
3. Investment and open ecology and talent training: Facing the whole world, we will continue to cooperate extensively with academics, industry and industry partners to create an open culture of artificial intelligence and cultivate artificial intelligence talents.
4. Solution enhancement: Introduce AI thinking and technology into existing products and services to achieve greater value and competitiveness.
5. Internal efficiency improvement: Application AI optimizes internal management, aligns with large-scale operation scenarios, and greatly improves internal operation efficiency and quality.
Among them, Huawei's full-stack full-scenario AI solution is the core of the entire strategy, including four aspects:
1. Serialized AI IP and chip Ascend based on a unified and scalable architecture.
2, chip operator library and highly automated operatorDevelopmentTool CANN.
3. MindSpore supporting end, edge, cloud independent and collaborative unified training and reasoning frameworks.
4. Provide full-process service ModelArts, layered API and application integration of pre-integrated solutions.
In this program, Huawei's self-developed AI chip is the most important. Xu Zhijun said, “If the progress of computing power is the main driving factor for the current development of AI, then the scarcity and cost of computing power are becoming the core factors that constrain the overall development of AI.”
The use of Ascend name in Huawei
When it comes to the name Ascend, Huawei first used it in the early days of smartphones. Huawei divided Ascend into four series D, P, G, and Y, which correspond to the flagship, high-end, mid-range, and entry.
In February 2013, Ascend P2 was released, carrying Huawei's own HiSilicon K3V2 quad core for the first time.processorBecause of the 4G license issue, P2 is not listed in China. After the release of Ascend D2, the same processor was used. K3V2 is the first successful mobile phone processor from Hisilicon.
In June 2013, Huawei directly skipped Ascend P3 and released the Ascend P6 mobile phone in London, England. This mobile phone uses the HiSilicon K3V2E quad-core processor, which plays a very important role in Huawei mobile phone products. It has become a major contributor to Huawei's mobile phone.
In May 2014, Huawei launched the Ascend P7 mobile phone in Paris, France, equipped with the HiSilicon Kirin910T quad-core processor. Until then, Huawei's own mobile phone processor split the Kirin series from Hisilicon.
In April 2015, Huawei launched the P8 new machine in London, England, equipped with the Kirin 930/Kirin 935 processor. At this time, Huawei began to slowly fade Ascend in the mobile phone, and directly named the mobile phone with the P series.
Five series and planning of Ascend chips
The current Ascend chip is part of Huawei's "Da Vinci Project" and is part of Huawei's full stack of artificial intelligence solutions.
Ascend is a series of AI IP and chips based on a unified and scalable architecture. The rising chips include five series of Max, Mini, Lite, Tiny and Nano, based on the "Da Vinci Architecture".
This time, two upgraded chips were released, namely Ascend 910 and Ascend 310.
Xu Zhijun said that the Shengteng 910 belongs to the Max series and is the single chip with the highest calculation density among all the chips currently released. The chip adopts a 7nm process and has a maximum power consumption of 350W. "The Shengteng 910 can reach 256 T. It is the largest AI chip with the largest number of single-chips released in the world, which is twice as high as the recent V100 of NVIDIA." Xu Zhijun said.
The Ascend 310 belongs to the Mini series and is currently the most powerful AI SoC for computing scenarios. The Shengteng 310 is the first product of the Shengteng Mini Series. It is said that this chip consumes 8 watts, adopts 12nm process, and can calculate up to 16TFLOPS. It integrates 16 channels of full HD.videodecoder.
In addition, Xu Zhijun also launched five AI products based on the Shengteng 310 chip, including the AI acceleration module Atlas 200, the AI acceleration card Atlas 300, the AI smart station Atlas 500, the AI all-in-one Atlas 800, and the mobile data center MDC 600.
The three series of chips, Lite, Tiny and Nano, will be released next year.
At the same time as the release of the chip, Huawei also released the Ascend cluster of large-scale distributed training system. In the design, the cluster will include 1024 Asced 910 chips with a computing power of 256P, which greatly exceeds the NVIDIA DGX2 and Google TPU clusters. This kindserverIt will also be launched in the second quarter of 2019 to help developers train models faster.
The following are the most concerned issues for Huawei AI chips:
Will the rising chip compete directly with Nvidia in the future?
The two chips released this time will be listed in the second quarter of 2019, but Xu Zhijun said in the subsequent media interview that Huawei's two AI chips will not be sold separately, but with AI acceleration module, AI server,cloud serviceThe form is for third party sales.
Xu Zhijun said, "We don't provide chips directly to third parties, but provide chip-based hardware and cloud services. We don't compete directly with pure chip vendors."
Why do you want to develop your own structure without using a partner program like Cambrian?
Why does Huawei, who has always been questioned by the outside world, build its own "Da Vinci Architecture" instead of the solutions of manufacturers such as Cambrian. Xu Zhijun replied in an interview, "Building a new architecture to support artificial intelligence chips is because Huawei is based on Huawei's understanding of artificial intelligence, and the demand for artificial intelligence is naturally generated based on the end-pipe cloud."
Xu Zhijun said that Huawei needs to cover the cloud, the edge, and the end to the Internet of Things, requiring a new architecture and a creative architecture. "The Cambrian is also very good, but can't support our full scene."
Huawei chief architect Party Wenshuo said that the Ascend chip adopts a unified DaVinci architecture: scalable computing, scalable memory, and scalable on-chip interconnect. Therefore, this is the world's first smart chip series covering the entire scene.
Why has Huawei been so focused on developing AI chips?
Xu Zhijun believes that under the current data privacy protection situation, many things cannot be done by the computing power on the cloud alone, and must be done on the end side. This is a very complex multi-objective optimization problem.
“This often faces the dual limitations of energy consumption and memory, and faces different needs in various scenarios. For example, in automotive applications, the response speed is very fast, and the processing accuracy of various pictures and videos is relatively high. In terms of sound, the requirement for noise reduction is very high. How to use GAN to separate voiceprints and content, which often involves personal privacy.” Xu Zhijun said that Huawei’s main purpose is to develop high-end aspects. The performance of the chip will be completed as much as possible on the end side, striving to provide the best user experience.
What is the relationship between Kirin chips and rising chips?
People close to Huawei said that Kirin chips will be the main mobile phone processor, and the rising chips are mainly used in conjunction with cloud services. Although chips such as the Shengteng 310 will be used in mobile phones, watches and other devices in the future, most of them need low power. The place to spend. In response, Huawei Fellow said that the relationship between the two is confidential and will be announced next year.