On the eve of the Shanghai Artificial Intelligence Conference (WAIC), the news of the "embargo" of high-end GPUs in NVIDIA's data center came. Suddenly, the wind was blowing.
The first time I entered the Expo Center, I asked about the causes and consequences of NVIDIA's "embargo".
After talking to a senior staff member of NVIDIA's China server OEM partner, I learned that the OEM had received the news a week before the big news broke. Nvidia's situation is not optimistic, the United States Internet companies are cutting orders, inventory pressure is large, I hope that Chinese partners quickly place orders. The explosion of big news objectively urges domestic buyers to quickly stock up on goods.
For some domestic Internet customers of NVIDIA Direct, if they cannot buy A100 directly, customers may buy A30 with half the performance of A100, and the price of two A30s is more expensive than one A100... If you want to buy a domestic alternative to the A100, it will take three to six months to complete the process just for the migration of algorithm software.
The performance of domestic GPU companies is uneven and difficult to say. But in the chaos, there is some hope.
Amid the clamor that "chip investment is dead", where is the future of China's technology industry? After research and interviews with many technology companies, this article intends to discuss three directions:
1.AI chip melee;
2. The robot wave is coming;
3. What should AR smart glasses do?
01
AI chip melee
Waic scene first impressed me, is the Flinthara showed the tongue moss testing Chinese medicine application, I went forward to spit out the tongue for ten seconds, and then answered more than a dozen multiple choice questions, and finally the screen gave a more versatile health advice:
Replenish qi and yang, and strengthen the qi and solidify the table.
Later, an engineer told me that Flinthara's products were not compatible with the CUDA framework.
This is somewhat surprising. There is a domestic strength to be incompatible with CUDA and can establish its own framework and ecology, I am afraid there is none. There are some technology companies, out of deliberately catering to the needs of certain specific markets and groups of people, for the so-called original and original, do not respect the basic law of the development of the technology market, and should be vigilant.
Pulling out the foundries and processes of the original flint products, combined with the market they face, is even more suspicious:
12nm from the American fab Grofond.
In stark contrast to the Flinthara is Denglin. The engineer introduced to me is very practical, he clearly said that it is not easy to sell chips alone, and the company provides more board solutions. The booth listed a combination of many third-party server OEMs, with their GPU chip CPUs, both Intel and Haiguang. The company focuses on cost-effective products, and the price of a single board is less than 10,000, which is half the price of the TCM application program displayed by Flinthara.
Denglin's engineers also told me that algorithm manufacturers and solution providers have a large right to choose products in the current industrial ecology, so the company has cooperated with a number of algorithm manufacturers.
Judging from the information obtained from the scene, denglin's market play is obviously very grounded, which makes full use of the mature market division of labor in the electronics industry, specializing in its own core links, unlike some domestic GPU manufacturers who are contracting local data center projects, although it looks like chips on the surface, but in fact, it is doing foot washing work, in order to do flowing water, it is gradually losing its strategic focus.
Of course, when it comes to domestic AI chips on the server side, a strong competitor in the current market is Huawei.
The staff of the server OEM Baode told me that in the Shanghai market, Huawei's Kunpeng and Ascend server solutions are very popular, and they are the most in shipments. I was surprised that Huawei actually had inventory... Huawei's chip shipments and inventory are like a mystery, although it is stuck in the neck, but it still maintains a strong sense of presence in the market.
In addition, the recent launch of the 2.5D architecture GPGPU using the 7nm process Chiplet process is also becoming a potential opponent of NVIDIA's high-end GPU in public performance indicators. Its chip solution is both practical and transcendent:
On the one hand, CUDA can be seamlessly migrated, and different solutions of boards, modules and cabinet solutions can be launched, winning the favor of server manufacturers such as Inspur; On the other hand, in BERT and ResNet50, the two most widely used core AI models in the field of computer vision (CV) and natural language understanding (NLP), the BR104 achieved the best performance over the NVIDIA A100, which is conducive to its full opening of the market.
AI chips, in addition to the above several schemes, there is actually another technical route: optical chips. This is a more cutting-edge field, the domestic Xizhi Technology has been in this field for a long time, its optical chip in a specific AI computing task greatly surpassed the GPU, commercialization is steadily advancing.
02
The robot wave is coming
There is a rather interesting phenomenon at the WAIC exhibition site, that is, the robot dog runs everywhere, and there are many robots on display.
I spoke with Wang Ji, an engineer on the robotics research team at the University of Shanghai for Science and Technology, and he revealed a major reason behind the popularity of robots: Boston Dynamics previously open-sourced code, providing a very convenient development foundation for domestic startups. Of course, because of Boston Dynamics' U.S. military background, the most critical code to solve the robot's instability is not open source. Therefore, most of the robot dog forms displayed on the WAIC site are relatively simple.
Let's start with the robot form.
The robot displayed by the research team of the University of Shanghai for Science and Technology uses an x86 motherboard with an FPGA chip, which is more distinctive. It is estimated that there are very few robot teams that will adjust FPGA chips on the spot.
Titanium Tiger robot did the demonstration, but the practicality is insufficient, I asked what the use is, they said it can be used for concierge hospitality, which is somewhat laughable.
However, The main business of Tihu is to design the servo motor contained in the robot trunk, which is then handed over to the foundry for production. The robot is full of body, using more than twenty motors, a single price of five thousand to eight thousand, the cost is not low. If the robot is truly commercialized, it is expected to bring a wave of development dividends to manufacturers who master the production capacity of servo motors.
Let's talk about robot dogs.
Judging from the on-site display, the robot dog is mostly used in scenarios with high danger such as power plant patrol and fire auxiliary. In addition to the general physical functions such as grabbing items, the robot dog also integrates a camera, which can build a high-precision map and send back the data. The machine dog solution in the depths of the cloud is based on Intel's industrial-grade CPU, which is installed with more than one piece, and there are huawei chip solutions used by other manufacturers. The price of a whole machine is in the hundreds of thousands of yuan, and the main customers are large enterprises and the government.
Objectively speaking, the functions of many robot dogs are similar, and it can even be said that the functions of some robotic arms are similar, but they take the form of dogs.
Under the market situation in which domestic manufacturers are mostly integrators aiming at specific applications, the robot and robot dog industry boom has benefited more and is the most certain supplier of key components such as chips and servo motors.
03
What should AR smart glasses do?
In addition to ai chips and robots on WAIC, AR smart glasses have also become a hot spot in the recent technology industry, especially the recent release of Nreal and Mijia AR glasses cameras, setting off a small wave of climax.
However, after trying out several products, I found that the current experience of AR smart glasses is difficult to call satisfactory.
The core crux of the matter is that the products on the market are basically "smart first and glasses later".
Rokid and Nreal's AR glasses are more like a head-mounted display, and the core function is to project video on in vitro devices. Even if it is a screen projection, the experience is not very good, and a friend told me that he has encountered the problem of layer fragmentation when watching movies many times... Not to mention that after putting on AR glasses, because the front of the line of sight is full of images, it is impossible to walk. In addition, there are fewer applications for AR adaptation, and some products can only be experienced on the web version.
Mijia's AR glasses can be worn to see the road clearly, its core function is mainly to take pictures, but it is a bit counterintuitive to use, changing the camera perspective sometimes requires hand breaking modules, and the texture of the film has some, which cannot be compared with mobile phones.
In general, according to the current price of a pair of AR smart glasses 2000 yuan ~ 2500 yuan, if it is only the main screen projection or taking pictures, the experience is only reluctant, it is difficult to imagine that it can detonate the consumer market. On the separate screen projection and camera functions, there are already very mature consumer electronic products, and the competition is very fierce.
So, is there any other way of thinking about AR smart glasses?
Recently, I visited an emerging smart glasses company formed by the core pioneers of the former flagship mobile phone, and the singularity is approaching, and I learned a different idea - "first glasses and then intelligence".
The founding team at Singularity Near told me that they recently did consumer research and found that the demand for entertainment audio and video (corresponding to screen projection) accounted for only 5% of smart glasses. About half of the respondents believe that the biggest use of smart glasses should be to better "see the world", such as accurately identifying objects, retrieving relevant information, navigating while driving, and so on.
Overall, smart glasses should make people better integrate into the physical world and increase productivity, rather than pulling everyone into the virtual metaverse like Meta.
Therefore, from this point of view, smart glasses should develop in the direction of "universal, out of the street, high frequency".
First of all, to be universal, do a good job in the basic functions of glasses, to ensure that users wear for a long time will not have a big burden, for this reason, the singularity of the near team has a special executive from treasure island glasses responsible for related professional work.
Secondly, to go out on the street, to be able to wear it, to be able to go on the road, and to have beauty, which is determined by the jewelry attributes of glasses. At present, many smart glasses are designed to be worn over the head, and light weight is a problem.
Finally, be high-frequency, load up high-frequency applications that consumers really need in the physical world that have unique value, rather than just lying in bed and watching the film.
It is from the fundamental needs of smart glasses that Singularity is close to gathering veterans from backbone enterprises in the electronics industry such as Huawei and DJI, gathering all kinds of talent backgrounds required by the smart glasses industry, reversing the realization path and key nodes of technology, and striving to make a different product that can detonate the market.
04
End
In the context of chip investment tending to be cold, the investment heat of robots and AR glasses has gradually increased. But if you disassemble it carefully, it is not difficult to find that whether it is a robot or AN glasses, like a chip, it is a complex system problem.
Complex systems mean that no matter which branch of the technology industry it develops, it is not a breakthrough that can be achieved by solving a single problem.
To build a complex system, the technology tree must be full, deep, and all in.
A technical shortcoming may be the key to the neck of the card for an industry, so it must be complete.
If a field is not dug deeply, it will not be able to achieve high gross profits, it will not be able to form competitive barriers, and others will be able to easily defeat it.
A core area without an all-in mentality, without investing enough strategic resources, ultimately can not succeed.
Only by pointing out, digging deep, and all in, the humiliation of the "card neck" of China's technology industry can disappear.