Shulou(Shulou.com)06/02 Report--

We know that computing power, algorithms and data are indispensable for AI landing. But there are priorities. If you ask AI service providers and application enterprises today what is the most urgent need for AI, I think the answer should be the same: computing power.

Last year, OpenAI released a global research report on supply and demand of AI computing power. The report shows that the demand for computing power in the artificial intelligence industry has shown an amazing growth trend since 2012. Demand doubles every three and a half months. In six years, the global thirst for AI computing power has increased 300000-fold.

And we know that the growth cycle of Moore's Law doubles in 18 months. In fact, the global AI special accelerated computing power, far from keeping up with the growth rate of demand.

There is no doubt that the hunger for math is the first wall on the road of AI development. And difficulties are also opportunities. In Huawei's technology and product system that fully embraces AI, there is a special department and its products that are aimed at the demand for AI special accelerated computing power.

In December last year, Huawei set up an intelligent computing business unit, which will provide AI computing support solutions for Huawei's AI system and the needs of the outside industry. In the product system of intelligent computing, along with x86 and ARM, Atlas artificial intelligence computing platform is a series of product solutions specially designed to meet the needs of AI computing power.

The name Atlas is the Hercules who propped up the universe in ancient Greek mythology. In today's story, let's enter the world of industrial AI and intelligent computing to see how Huawei's Atlas supports the origin of all the stories of AI computing.

The bottleneck of arithmetic: the first Wall on the Road of AI

We know that AI, like steam engine and internal combustion engine, is a general technology, which can be combined with the production system of various industries to explore the possibility of further improvement in productivity and production efficiency. So the AI calculation power is the coal and oil in this productivity upgrade-without the power to "burn", everything will not work.

In fact, the problem of calculating bottleneck is not that simple. Not as long as you have the chip, you can do everything. In view of the bottleneck of computing power in industrial practice, there are at least three practical problems that need to be solved:

1. Due to insufficient computing power, the enterprise AI business cannot be started at all.

Today, enterprises need plenty of AI computing power. Without the special acceleration power of neural network as a support, enterprises will face the problems of long training time and excessive energy consumption when applying AI, resulting in a lot of industrial practice can not be carried out at all.

We once interviewed a doctor in the field of bioengineering, whose AI algorithm was originally well designed, but because the school lacked the relevant computing resources, it took several months to train at a time, resulting in the research project unable to meet the time limit at all. There are so many cases that if AI wants to run in practice, the supply of arithmetic is a prerequisite.

2. Edge scenarios cannot be infiltrated, and AI is lack of practical operation.

Another problem with AI computing is that computing power is delivered through cloud services, so that all tasks have to be sent back to the cloud. The direct problem brought by this is the lack of practical possibility for many tasks that require low latency and small data training.

For example, a large number of intelligent cameras are used in intelligent transportation, and if all the camera data are sent back to the cloud for processing and then making decisions, then it is very likely that the traffic situation has changed: it cannot play the role of real-time monitoring and real-time analysis.

3. Lack of vertical solution, the industry does not know how to apply AI

There is also a situation that is specific to a particular industry and requires a relatively complex arithmetic solution. It is not the transfer of the basic AI capability, but the combination with the industrial production logic and the existing technology system. Then this situation can not just simply quote the computing power, but need to transform and customize the computing power to meet the final needs of the enterprise.

Such a relatively complex arithmetic demand is particularly significant in energy, automotive and other industries. Computing power providers must work with industry vertical technology solution providers and data companies to build complete industry vertical computing solutions.

The combination of the three forms a situation of widespread lack of computing power and complex demand for AI today. The first answer to these questions for Huawei's intelligent computing business is to send a "war hero" named Atlas.

Hercules holds up the canopy: the intellectual Storm brought by Atlas

The name Atlas is the son of Iapatos in Greek mythology and the brother of Prometheus the fire robber. Because of his great power, he supported the whole sky and universe in the western pole of the world. So it is called Hercules or God of Heaven.

And when the name appears in the market as an intelligent computing platform, it is clear that Huawei's hope for it is to prop up the sky of the era of everything intelligence.

What on earth is Atlas?

Last October, Huawei released a series of AI chips based on Ascend Teng, as well as a full-stack and full-scene AI solution.

(Huawei full stack full scene AI architecture)

These capabilities and the mainstream heterogeneous computing components in the industry, integrated in various specifications of computing acceleration products, constitute the Atlas artificial intelligence computing platform that Huawei brings to the industrial world today.

Currently, Atlas has an Atlas G series of products that provide AI computing power upgrades to CVM, and has launched end-to-side Atlas 200 AI acceleration modules, Atlas 300 AI acceleration cards to support the industry's highest density 64-way video inference, Atlas 500 smart stations for edge side, and Atlas 800AI all-in-one machine targeted at the enterprise domain's one-stop AI platform.

These products are like Lego bricks, covering a variety of cloud scenarios from large to small, and can connect with Huawei cloud services, and build a complete industrial AI computing practice solution based on a variety of permutations and combinations.

Let's see what every member of the Hercules Atlas family can do.

The Atlas 200 AI acceleration module, which is only the size of two coins, can be cleverly integrated into a variety of terminal devices, increasing the AI computing power of a large number of terminal devices by seven times. For example, in drones, cameras, service robots. Atlas 200can give these devices the ability to independently complete face recognition, posture recognition and other tasks, and the data does not have to be sent back to the cloud to complete the operation immediately. At the same time, it is not afraid of the embarrassment of offline equipment.

(Atlas 200)

The Atlas 300 AI accelerator card is a machine vision task acceleration device for data center and edge-side server scenarios. Support a variety of data precision, a single card can provide 64TOPS INT8 computing performance. Through Atlas 300empowering, camera, face recognition and other machine vision equipment, we can quickly and accurately support the access and forwarding of multi-passers-by face video stream and picture stream, as well as portrait integrated services. We only need to use a qualified face picture to complete the search and calculation of the target identity information, which makes the scene such as face recognition fast, accurate and stable.

(Atlas 300)

The Atlas 500 Smart Station, the industry's leading edge product with integrated AI processing power, is only the size of a set-top box, but can achieve 16-way HD video processing capacity and supports outdoor environments from-40 ℃ to 70 ℃. With Atlas 500s and cameras, you can easily complete a single scene, such as intelligent monitoring of hospitals, schools and homes for the elderly.

(Atlas 500)

The Atlas 800AI all-in-one machine is an integrated computing power supply equipment used to build a telecom service AI model training platform. It is a fully modular design and supports up to 32 Atlas AI acceleration cards. Through it, enterprises can easily complete AI task training and deployment, and greatly improve the ability of AI business understanding and application.

(Atlas 800)

As you can see, Atlas's complete product system is characterized by the ability to popularize AI computing power to all scenarios. Whether it is outdoor, offline, or combined with the old technology system, it can be easily done by using the arrangement and combination of Atlas. Integrating complex things into its own system and leaving simplicity to application scenarios, Atlas is very important to the popularity of AI.

Such a series of systematic intelligent computing products, as well as the platform services behind, Atlas has been widely used in smart cities, intelligent transportation, intelligent medical, enterprise-level AI reasoning and other fields.

It can be said that a storm about intelligent computing is brewing in the new opportunities slowly opened up by Atlas.

The Application key of Wanye AI: from Atlas platform to Intelligent Computing Ecology

Going back to the bottleneck problem of AI computing power that we began to discuss, we can find that the marketization and production of Atlas first solved the embarrassment of enterprises' lack of AI computing power and had to rely solely on importing AI computing products. Then, through flexible and changeable, the product form of various scenarios of end-to-side cloud is realized, and the AI computing power requirements of end-side and edge-side are solved.

Then the next step of capacity completion, the focus is on the construction of Atlas ecology. Only by building an ecological and industrial alliance and combining the industry AI solution providers with the underlying AI computing power provided by Atlas, can AI have a real application key to enter thousands of industries.

To this end, Huawei Intelligent Computing Industry Summit was held at the just-passed Huawei China Ecological Partnership Conference 2019, at which Huawei intelligent computing and industrial technology experts, financial, Internet, telecom operators and other industry customers, Yitu, Issa, Huayu, Geling Deep pupil and other industrial partners jointly established Huawei Intelligent Computing Ecological Alliance.

The important task of such an alliance is to improve and develop the Atlas industrial ecology. On the one hand, partners can use Atlas to complete their own AI technological innovation and industrial AI solutions, on the other hand, customers can accurately find customized and upgraded AI solutions in the Atlas ecology, so as to be responsible for industry characteristics and industrial upgrading needs.

The industrial value of Atlas lies in the implementation of the characteristics of AI computing power on IT infrastructure. Atlas constructs the whole frame structure of ecology, from the bottom computing power, middleware to the upper application. With this new infrastructure, AI and industrial customers can find a new docking stage and make a lot of ideas that can't be landed come true.

The next step of ecological convergence is the continuous expansion of industry applications, accompanied by the further iteration of technological solutions.

According to the current case data, Atlas artificial intelligence computing platform has completed the landing application of scenarios in many industries, and can infiltrate different combinations of scenarios on the cloud side:

For example, in the case of a bank, Atlas edge computing is used to quickly complete the iteration of face recognition instead of swiping card in the bank scene, with an identification accuracy of more than 95%, and non-stop traffic can be achieved.

In the traffic case of a city in the south, Atlas has been widely used in all aspects of intelligent traffic monitoring equipment and intelligent brain, so that the complex high-mobility traffic scene has been fully systematic AI computing power. To achieve a wide range of human, vehicle, certificate data collection, and achieve the data collision, multi-rail integration, multi-dimensional control and other functions. In the application area, the traffic light is adjusted according to the congestion through the intelligent transportation system, the road speed is increased by 9%, and the congestion time during rush hour is reduced by 15%.

In cloud-side computing scenarios, the effect of Atlas is also obvious. For example, a genetic biotech company uses Atlas-based accelerated cloud services to greatly shorten the time required for gene sequencing. Compared with traditional x86-based cloud service platforms, Atlas helps enterprises improve computing efficiency by five times and reduce end-to-end comprehensive costs by 30 per cent.

It can be seen that with the deepening of the understanding of the process of generalization of AI, the industrial foundation and application practice of Atlas are superimposed step by step. Through the penetration of cloud-side multi-scene applications, and the effective acceleration of vision, voice, data intelligence and other AI tasks, together with the construction of industrial ecology, Atlas has in fact become the only hardware platform product in the industry that can provide panoramic AI solutions.

Maybe today is when the God of Optimus just started to work. Predictably, however, Huawei's goal of creating a full-stack and full-scene AI industry ecosystem is clear. The cloud-edge integrated intelligent computing layout, as well as the evolution of AI in various industries based on Huawei intelligent computing, all need the pillar of Atlas as the foundation.

With the passage of time, AI, like all technology buildings, was not built in a day. And Atlas is a basic project about the future that cannot be bypassed today.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.