Shulou(Shulou.com)06/03 Report--

I've had a lot of interviews with AI companies. But what often impresses me is those enterprises that "want to use AI but do not succeed". For example, last year, I interviewed a logistics company in Hunan. They hope to install a smart camera system in their storage park, use AI to help monitor cigarette butts, smoke and other unexpected situations, as well as regional face recognition to help security.

We have learned that although the original intention of this design is very good, there is a situation in practical application-with ordinary cameras, the AI algorithm is too inefficient; and the smart cameras on the market cannot meet corporate expectations; developers want to buy AI acceleration cards to help the camera system deploy algorithms, but find that mainstream acceleration cards are not only very expensive, but are often not in stock, and they have to wait a long time.

Thus it can be seen that although the third upsurge of AI is greatly stimulating the nerves of enterprises, the door of AI is still "sealed" for the majority of enterprises. What's the problem?

Think calmly that AI needs not only cloud computing and big data services, but also the deployment of large-scale computing power in edge scenarios and end-to-side scenarios, and the customized demand for AI computing power in specific production scenarios, as well as the cost load capacity of the enterprise. For example, this logistics company, they must get a lot of computing power in the camera and campus edge computing scenarios, and the cost can not be too high, in order to really build a smart park in mind.

However, the ideal always collides with the reality of bony feeling. The rapid outbreak of the AI industry has led to a sharp increase in the demand for side-to-side computing power, but it is actually encountered that only a few suppliers in the industry can provide related products. For startups, small and medium-sized enterprises and real economy enterprises eager for intelligent transformation, the hunger and thirst for side-to-side computing power is becoming a real pain.

Today, this widespread anxiety is also becoming a driver of change in the intelligent computing market.

Scorching drought: the arithmetic Dilemma of the Eve of "Intelligence +"

In essence, the AI computing problem is not a single technology or contradiction between cost and performance, but a complex set of comprehensive problems. Such a tangled dilemma for enterprises eager for AI is like the legend in ancient times that there are ten days in the sky, and ten suns bake companies and developers who have just entered the AI industry from different angles, often leaving them at a loss.

The plight of the enterprise park we interviewed is very representative: when we thought of buying end-side AI acceleration products, we were surprised to find that the cost of hold could not live, or even could not be bought at all, which is the collective pain point for a large number of small and medium-sized enterprises today.

This is determined by market supply and demand. In the very important field of end-to-side AI computing acceleration, the product options available on the market are indeed very rare.

At present, the main supplier is basically Nvidia. As an important part of Nvidia's AI strategy in recent years, the embedded intelligent products of to B have become the core part of its strategy. Products such as TX2 embedded modules and P4 accelerator cards have occupied the mainstream, and the new Jetson Nano products just released at the GTC conference in March have also attracted wide attention.

Intel has also begun to enter this field in the past two years. After acquiring Movidius at the end of 2016, Intel gradually launched a series of Movidius products for AI and machine vision scenes, but its voice is relatively low in the market. AI startups such as Cambrian and Bitland have also launched related products, but they are not yet mature.

The dominance of a large company and the scarcity of product choices have led to a long-term imbalance between supply and demand in this market, hampering the hands of enterprise users and developers.

For the consideration of computer control, in recent years, international Internet giants such as Amazon, Google, Facebook, including domestic giants such as BAT, have begun to lay out artificial intelligence infrastructure construction, such as AI chips. However, "deep-pocketed" Internet bosses may be able to achieve self-redemption, so small and medium-sized enterprises that need to complete industrial upgrading and intelligent transformation with the help of AI, how to solve the math dilemma?

As previously analyzed, if small and medium-sized enterprises want to use side-to-side AI computing power today, there are several hurdles that must be crossed: computing products are too expensive to support large-scale deployment; the purchase process is too long, and official websites often have to queue for a long time, resulting in some enterprises having to choose consumer products at the top. Poor compatibility and lack of coordination at the edge of the cloud lead to difficulties in deployment and development-which in turn lead to the need for help in solving the difficulties, and the high threshold for technical personnel needed. creating a development dilemma-- it's a bit like a vicious circle.

Under such circumstances, if there is always a lack of new "breakers", it will lead to a long-term mathematical dilemma. What enterprises need is not only the rare "well water" or short-term "shade" under the scorching sun, but also Hou Yi, who can "break the hot day from the root": break through from the AI infrastructure level and use performance-to-price ratio and product advantages to break these rules.

Casting Arrow: the Atlas of "counter-common sense"

The insufficient supply and rigid demand of side-to-side AI computing products make the breakers of this market look very precious today. The good news is that enterprise users and developers don't have to wait long.

At the Shenzhen Station of Huawei Intelligent Computing Conference on April 10, Huawei officially launched the Atlas artificial intelligence computing platform, which is also an important product of Huawei's intelligent computing family. Four products including Atlas 200 AI acceleration module, Atlas 300 AI acceleration card, Atlas 200 DK AI developer kit and Atlas 500 smart station were unveiled at the conference. The common feature of these four products is that they work in side-to-side scenarios that are in urgent need of computing product solutions, and match the mainstream performance with "quite lethal" performance-to-price ratio.

First, let's review these four products:

Atlas 200is an end-to-side AI acceleration module for cameras, drones, robots and other hardware, which can handle real-time analysis of 16-channel HD video.

At the same time, in order to help developers develop in the Atlas 200environment, Huawei has also launched a matching Atlas 200DK AI developer kit.

The Atlas 300AI accelerator card can provide AI acceleration in the fields of video analysis, OCR, speech recognition, precision marketing and medical image analysis, helping developers improve the computing performance of video analysis and high-density reasoning scenarios.

The final Atlas 500 smart station is an original edge smart station product that can deploy huge AI computing power on the premise of low power consumption, filling the relative gap in the edge-side AI acceleration industry.

It can be seen that these four AI computing products are closely related to the machine vision and computational reasoning capabilities encountered by enterprises in deploying AI in actual scenarios, and they are also the mainstream products that a large number of developers actually need today.

At the same day's press conference, the most striking point is that after the release of these four products, Huawei disclosed the prices of the four products on the spot. Friends who have knowledge of to B business should know that the press conference in this field generally does not mention the price. And the behavior at the Atlas press conference is really a bit of a "counter-sense". On the one hand, it is inconsistent with everyone's impression that China is a low-key style, and at the same time, it is not in line with the industry law of to B industry.

Why is Huawei willing to "post the price" unusually on the spot? Looking at the logic behind it, we can guess that there are two reasons: one is that Huawei does have the strength to have absolute confidence in product performance and value for money; the other is that Huawei has seen the price transparency of computing power, which is of far-reaching significance in the AI computing industry at this stage-if it makes sense, then Huawei is willing to act as a subverter.

First of all, let's take a look at Huawei's "strength". We can visually compare Huawei's Atlas with the popular Nvidia mainstream products through the core parameters and prices provided by the government. It is not difficult to see that Huawei Atlas does sell edge AI computing products at seductive prices to developers and small and medium-sized enterprises on the basis of stronger mainstream performance and computing power-Atlas provides several times the computing power of mainstream products at similar prices.

The prices of Nvidia products vary from agent to agent. The comparison of data is based on official data, and the actual application effect depends on the specific judgment of business and scene environment)

At the same time, it is worth noting that Huawei has always been known for its strong engineering capacity, quality service and abundant supply. So the invisible competitiveness of Huawei Atlas lies in reducing the time cost and unexpected cost of developers. It also reflects Huawei's strategic intention to build a smart world of the Internet of everything: to let all industries really enter the inclusive AI model, so that they can use it well, affordable and at ease, and have already thought of reducing costs at the time of casting arrows.

The Atlas product line essentially achieves three cost reductions in mainstream AI application scenarios:

Direct computing cost: compared with Nvidia's mainstream products, it greatly reduces the unit computing price and solves the pain points that users are most concerned about.

Future industry cycle cost: Atlas's product design idea is to integrate big computing power into a single component. This is in line with the increasing complexity of the AI algorithm, large-scale parallel deployment and seamless computing in the cloud, which will continue to increase the test of the computing power of a single acceleration module in the future, and also reduce the replacement cost after the future computing demand.

Development cost: because Atlas products are based on the production of Penton 310 chips and have the advantage of full-stack development, an architecture allows cloud-side full-scene deployment of AI, reducing compatibility costs. This is actually the area that AI developers are most concerned about today.

A "anti-common sense" product release made Atlasas attract attention with high profile and high competitiveness as soon as it entered the market. The logic behind this is that when AI's thirst for numeracy has become the norm, then what Atlas must do is break through the rules and become a subvert of the "normal".

Perhaps it can be understood that every Atlas product is a sharp arrow forged by Huawei in order to pierce the "math sun".

Sheri: releasing the Force from Product to Ecology

At the level of embedded AI and edge intelligence, the long-standing status quo is that developers have to make do, and enterprise users are confused. When everyone thought this would be a long-term reality, Atlas was like a stone falling into the lake, causing ripples in the dreary industrial shape and driving a change in both directions.

On the first level, the landing of Atlas products accelerates the industrial change of the whole stack and scene.

First of all, as a powerful product, Atlas enriches Huawei's intelligent computing sequence and lays a solid foundation for the landing of a large number of AI services and the ecological generation of AI. At the same time, the direct landing of the products based on the Teng 310 chip also triggered the expectation of the environment for the Teng 910 products, as well as the expectation of more Teng series chips. From this point of view, Atlas products are very malleable platform products. A new track has been launched for Huawei in the relatively blank market area of side-to-side AI acceleration products.

On the other hand, Atlas has a more far-reaching impact on the ecology of AI industry.

While attacking the pain point of the industry, its "explicit real price" play is likely to force the industry into a universal cycle of falling costs, thus prompting the industry to break the potential monopoly trend, re-unify the track, and catalyze the release of new computing cost standards, which in turn may lead to the reconstruction of AI infrastructure and the great improvement of development imagination.

Whether it is Huawei's internal product significance to strategic value, or the impact on the entire industrial ecology, Atlas is affecting the future trajectory of the AI industry with technological value and product logic. And the two clues eventually converge into one sentence: AI can achieve true inclusion through innovation at the infrastructure level.

The starting point of Pratt & Whitney AI is the decline in computing costs, so what's the next step? Logically speaking, the next responsibility that Atlas needs to undertake is to further empower developers, open up the ecological situation of AI development, and then find a way for small and medium-sized enterprises and start-ups to enter the "smart +" world.

This step has indeed been taken. A month after the release of Atlas products, Huawei will officially launch the "Huawei Atlas artificial Intelligence developers Competition" at the Huawei Intelligent Computing Conference in Suzhou on May 10. For more information, please follow "Huawei Intelligent Computing" WeChat Mini Programs.

It is worth noting that in addition to the four released Atlas products, the contest also includes the Atlas800 deep learning system, which should be a heterogeneous server that supports AI cloud training. With the addition of this sharp weapon, Huawei has actually completed the layout of AI training and reasoning products that can be deployed in the cloud.

According to the rules of the competition, participants can rely on the Atlas artificial intelligence computing platform to create software and hardware solutions for different scenarios. For example, smart cameras, drones, robots, smart hardware, edge AI hardware based on Atlas 200 AI acceleration module, or face recognition, vehicle recognition, image recognition solutions based on Atlas 300accelerator card. Interested developer friends might as well log on to the "Huawei Cloud" website, search for "Huawei Cloud Competition" and choose "Atlas track" to see if they have the opportunity to show their skills.

Reviewing the logic of the full text, there are several key points that may be worthy of the attention of enterprise users and AI developers:

1. Side-to-side AI computing products are very important, but the market contradiction is quite prominent.

2. The entry of Huawei Atlas series products means that the default situation of this market has changed. High performance-to-price ratio and ease of use are likely to become the main competition points in this market in the future.

3. Pratt & Whitney AI is bound to make more contributions to AI ecology from the beginning of Computational Pratt. The construction of AI infrastructure still has a long way to go.

The good news is that even if the pain of corporate AI seals is on fire, the "Arrow of the Sun" by Atlas and countless AI developers has left the bowstring.

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.