Recently, Ren Zhengfei mentioned a point in his public speech: "5G is only a pediatrics, and artificial intelligence is the strategic location for Huawei's development."

Today, Huawei launched a charge against this strategic location, officially released the "world's most powerful AI processor rising 910", and the AI ​​open source computing framework MindSpore.

AI Big Mac

The chip industry has suddenly become popular "big" recently. California AI startup Cerebras Systems announced the world's largest chip, the chip called "The Cerebras Wafer Scale Engine" has 1.2 trillion transistors, a little larger than the standard iPad.

Although the launch of Huawei's Shengteng 910 is not so exaggerated, it is still a big guy. Its half-precision (FP16) is 256 TeraFLOPS, the power is twice as high as the NVIDIA Tesla V100 (125 for the NVIDIA Tesla V100), and the integer precision (INT8) is an amazing 512TeraOPS. At the same time, a 128-channel full HD video decoder has been added to the chip.

With such a strong calculation, power control is very good. The power consumption of the Shengteng 910 to reach the specification power is only 310W, which is significantly lower than the 350W of the design specification. This may be due to the fact that the chip uses a 7nm enhanced EUV process, and the single Die has 32 built-in DaVinci cores.

The DaVinci core, or Da Vinci architecture, is Huawei's self-developed new computing architecture for AI computing features, featuring high computing power, energy efficiency, and flexible cutting. Specifically, the DaVinci architecture uses 3D Cube to accelerate matrix operations, greatly increasing the AI ​​power per unit of power consumption. Each AI Core can achieve 4096 MAC operations in one clock cycle, compared to traditional CPU and The GPU achieves an order of magnitude improvement.

At the same time, in order to improve the completeness of AI calculation and the computational efficiency of different scenes, DaVinci architecture also integrates various computing units such as vector, scalar and hardware accelerator. At the same time, it supports multiple precision calculations, supports the data accuracy requirements of the two scenarios of training and reasoning, and realizes the coverage of the full scene of AI.

At the 2018 All-Link Conference, Huawei's AI Chip Rising 310 was the debut of the Da Vinci architecture. Among them, Da Vinci Core is only a part of NPU. Da Vinci Core is also subdivided into many units, including core 3D Cube, Vector vector calculation unit, Scalar scalar calculation unit, etc., which are responsible for different computing tasks to achieve parallel computing. Models, together to ensure efficient processing of AI calculations.

As the strongest member of the Da Vinci architecture, the actual performance of the Shengteng 910 is excellent. Xu Zhijun said that the Shengteng 910 has been used for actual AI training tasks. It works with MindSpore and shows nearly double the performance improvement compared with the existing mainstream training single card with TensorFlow. The number of pictures trained per second has increased from 965 to 1,802.

"The overall technical performance of the Shengteng 910 exceeded expectations, as the most powerful AI processor, deserved!" Xu Zhijun said.

In addition to the protagonist Shengteng 910, another product announced at the Global Connection Conference last year has also received new news. According to Xu Zhijun, MDC based on the Shengteng 310 and many domestic and foreign mainstream car companies have already cooperated in the park bus, new energy vehicles, and automatic driving. The Atlas series of boards and servers based on the Shengteng 310 and dozens of partners have settled industry solutions in dozens of industries such as smart transportation and smart power. Huawei Cloud also provides cloud services such as image analysis services, OCR services, and video intelligent analysis services based on the Shengteng 310.

The Shengteng 910 and the Shengteng 310 will be the beginning. Huawei will also launch a series of AI processors. For example, the Soaring 920 for training scenes will be launched in 2021, and the edge computing scene will be 320.

AI processor best friend

The AI ​​processor must be combined with a computing framework to deliver a powerful firepower. MindSpore is the best friend of the Shengteng series.

The industry's most popular AI computing framework includes TensorFlow, PyTorch, PaddlePaddle, etc. Now there is MindSpore. It is a unified training and inference framework that supports end, edge, cloud independence and collaboration, and can support the full scenario proposed by Huawei. Huawei hopes to achieve one-time operator development, consistent development and debugging experience through this complete software stack, in order to help developers achieve one-time development, application smooth migration capability on all devices, edges and clouds.

The MindSpore framework supports all devices from large to small, while also supporting local AI calculations for privacy protection. The data that this framework delivers to the cloud can be processed gradients and model information without privacy information, rather than the data itself, so as to achieve cross-scenario collaboration under the premise of ensuring user privacy data protection. In addition to privacy protection, MindSpore will also protect the model's security-trusted design by protecting the Built-in into the AI ​​framework.

MindSpore's development philosophy is the AI ​​algorithm, code, which makes the development state more friendly and significantly reduces model development time. Taking a typical NLP (natural language processing) network as an example, MindSpore can reduce the core code by 20% compared to other frameworks, and the development threshold is greatly reduced, and the overall efficiency is increased by more than 50%.

MindSpore can not only support the rising processor, but also the CPU and GPU. In master-slave control mode, CPU and GPU interaction introduce memory and data overhead. MindSpore performs all control and execution of neural network model training on the chip, reducing the interaction time with the host CPU and making it faster.

MindSpore provides data scientists and researchers with new tools to make theoretical exploration and innovation easier and more efficient. In order to better promote the application of AI, Xu Zhijun announced that MindSpore will open source in Q1 in 2020 to help every developer.

Huawei AI burst

Huawei began investing in AI foundations and algorithms in 2012, but for the public, its research progress and results are mysterious. Until December 2016, Huawei Glory released the glory Magic, the first time in the industry to introduce artificial intelligence systems into mobile phones, so that AI results surfaced.

On October 9, 2018, Huawei released its full-stack full-scenario AI solution at the Joint Conference 2018, covering a multi-layer solution from the terminal to the cloud, from the AI ​​chip to the deep learning training deployment framework, officially marking the Huawei AI program. Full launch.

As the global technology giant, Google and Facebook have been in the AI ​​field for a long time. Google first launched the AI-specific chip TPU, and Facebook deployed AI technology on all of its social software. Although Huawei’s aggressive entry into AI is late, the overall strategy is complete:

· Fundamental research in investment: building efficient data (less data requirements), efficient energy consumption (lower computing power and energy consumption), safe and reliable, automatic autonomous machines in areas such as computational vision, natural language processing, and decision-making reasoning Learn basic skills.

· Create a full-stack solution: 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.

· Investment openness and talent development: Globally, continue to work extensively with academic, industry and industry partners.

· Solution enhancement: Bring AI thinking and technology to existing products and services to achieve greater value and greater competitiveness.

· Internal efficiency improvement: Apply AI to optimize internal management, aim at massive operation scenarios, and greatly improve internal operation efficiency and quality.

The processor chip and computing architecture released today is the concrete landing of Huawei's AI strategy, which laid a good foundation for Huawei's future AI development.

Although Huawei is facing many challenges, it is still full of confidence. Xu Zhijun said that the external events have no effect on the development of Huawei AI, and everything is currently developing according to plan. With the advent of the 5G era, Huawei hopes to ensure that enterprise AI applications are always in the best state by implementing closed-loop and real-time updates of models, and AI technology with 5G, cloud, Internet of Things, edge computing, blockchain, big data, Technologies such as databases are fully coordinated to deliver greater value.