Home > News content

HUAWEI AI strategy complete disclosure! AI chip exposing punch TPU compete for NVIDIA

via:博客园     time:2018/10/10 19:48:30     readed:743

Wen / Anne, chestnut, Qian Ming

Source: qubits (QbitAI)

If you don't sound, you will.

Just now, HUAWEI AI strategy has been fully disclosed.

At the 2018 Huawei All-Connectivity Conference, Huawei's rotating chairman, Xu Zhijun, announced at one breath high density: Huawei's AI strategy, Huawei's AI all-stack all-scene solution, Huawei's own development of two AI chips with a unified Da Vinci architecture.

Last year, HUAWEI, which had lost its mission vision because of AI, is now fully engaged in the battle to enter AI.


AI strategy for the first time

After today, some people say that HUAWEI does not have AI. Maybe Xu will be unhappy. For AI, HUAWEI group's vision has changed.

At the end of last year, Huawei decided on its new vision and mission: the intelligent world in which all things are interconnected. In order to highlight intelligence, the theme of the conference was set as

In the first 20 minutes of the conference, Xu pointed out a truth: AI is too important.


As a universal technology, the impact and value of artificial intelligence on the future is not to be underestimated, but also the key to building future competitiveness, Xu said.


At the scene, Xu announced the AI strategy for HUAWEI for the first time:

Investment basic research: realizing safe, reliable, and self-made machine learning basic ability.

Create Full Stack Solutions: Create Full Stack Solutions for Cloud, Edge and End Scenes, Independent and Collaborative, Provide Abundant and Economical Manpower Resources, Easy to Use, High Efficiency, Full Process AI Platform

Investing in open ecology and personnel training, facing global partnership, building an open ecosystem and cultivating talents.

Solution enhancement: introducing AI thinking and technology into existing products and services

Internal efficiency improvement: use AI to optimize internal management.

Full stack full scene AI solution

Next, Xu Zhijun released Huawei full-stack full-scenario AI solution, which includes data acquisition, training, deployment and other links within their own framework, the main purpose is to improve efficiency, so that AI application development more easy and convenient.


The full scenario includes five scenarios: Consumer Device, Public Cloud, Private Cloud, Edge Computing, and Industrial IoT Device.

The emphasis is on the whole stack, which consists of four parts.

One is Ascend (Sheng Teng), AI IP and chip, all based on the Da Vinci architecture. The chip is divided into 5 series, Max, Lite, Mini, Tiny and Nano.

The other is CANN (Compute Architecture for Neural Networks), which is a highly automated operator development tool.

According to official data, CANN can increase efficiency by 3 times. In addition to efficiency, we also take into account the performance of operators to adapt to the rapid development of academic and industrial applications.


Third, the MindSpore architecture, which integrates training and reasoning in a friendly way, integrates various mainstream frameworks (independent and collaborative):

Including TensorFlow, PyTorch, PaddlePaddle, Keras, ONNX, Caffe, Caffe 2, MXNet and so on.

This architecture fully adapts to the scene of end, edge and cloud.

Fourth, ModelArts, a machine learning PaaS (Platform as a Service), provides full process services, hierarchical APIs, and pre-integration solutions. It is used to meet the different needs of different developers and promote the application of AI.


Previously, Huawei had released two sets of solutions, Huawei Cloud EI for government, enterprise and HiAI for intelligent terminals. The full stack full scene AI solution released today will provide more perfect support for both.

Rising 910: single chip with the highest density


Xu Zhijun said:


This is a Max series of Shengten910, called by Xu Zhijun is

The performance data of the lift 910 is very strong. The half precision is (FP 16): 256 Tera FLOPS, integer precision (INT 8): 512 Tera FLOPS, 128 channel full HD video decoder - H.264/265.

What is the performance of the chip? HUAWEI compared with friends. The battles included Google TPU v2, Google TPU v3, Nvidia V100 and Huawei's Rising 910.


What would happen if we set up 1024 sets of the Sun Teng 910? Xu Zhijun said that there would be

This large-scale distributed training system is called

But with all this talk, what does Mount Lu look like after 910? Xu Zhijun said that 910 of the market will be launched in the second quarter of next year on HUAWEI cloud. So if you want to have a look at the flowers, you will have to see you next year.

I wonder if NVIDIA listened.

Rising 310: efficient computation of low power AI SoC

Thinking that 910 of HUAWEI's deployment is the whole thing? Naive, Xu Zhijun brought with us a chip on the edge device.

There are quite a few members of the rising series at the edge end, according to the power consumption from small to large, the four chip models are Nano,Tiny,Lite and Mini., respectively. Find four English descriptions in one breath


Among them, the three models of Nano,Tiny,Lite are all today


Compared with the Rising 910, the edge series of Rising Chips are much more user-friendly, smart phones, smart accessories, smart watches and other edge devices, are the edge series of Rising Chips shelter. Follow up, HUAWEI will also launch a series of AI products.


OMT: HUAWEI friends circle

Before, there have been rumors that Microsoft will adopt HUAWEI's latest AI chip. Today, Microsoft global senior vice president and Microsoft Asia Pacific R & D group chairman Hong Xiaowen is present at the conference site.

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments