Address: Chuangke Town, Zhongguancun, Hot Spring Town, Haidian District
Copyright © Zhongguancun Chuangke Town (Beijing) Technology Co., Ltd. 京ICP备16045011号-2
The Cambrian Unicorn has become a variable, but the Chinese AI chips are still ahead of us.
As an important era of life and species explosion, "Cambrian" has a unique meaning. Cambrian is already the AI chip solution supplier of AI Unicorn in China, and undoubtedly expects to play a pioneer role in China's AI industry: to promote the development of China's AI industry in the dual identity of architect and ecological leader.
Admittedly, China's AI industry has grown rapidly in the past few years, and Cambrian has played a very important role in it. However, Cambrian market cultivation performance is not as bright as its valuation. First of all, its terminal customers have not been successfully expanded. Although Cambrian officials have said that many customers are considering or have introduced its architecture, in fact, the only terminal customer in the Cambrian market so far is Huawei.
Secondly, the Cambrian also launched the cloud-end scheme, which is not only IP, but also entity computing card. This approach is like Arm launching its own mobile phone chip, but also telling the market that to use Cambrian IP to make chips, the first opponent that may encounter is Cambrian. In this case, in addition to the cooperation of Sino-Tech Dawning, there is no new customer information. Even Huawei, the original terminal customer, has to switch to its own uplift plan to completely get rid of the Cambrian.
Huawei's AI development strategy emerged at the Huawei All-Connection Conference just held on October 10. It not only completely cut off from the Cambrian, but also directly aimed at the Cambrian market.
So, is the Cambrian really revolutionary to AI industry in the end, or will it only end up as a "merciless tide sending home"?
There is no doubt that the Cambrian has proved its technical strength, but there is still not enough technology. What is the industrial status of the Cambrian and the future development of the Cambrian? From the current situation of AI industry, it is not optimistic.
Cooperation with Huawei is full of paradox
As the only listed Cambrian related architecture, Cambrian Kirin chips in Huawei have two generations of products. They play an important role in promoting AI ecology, whether in Kirin 970 or the latest Kirin 980. However, as the core function of two generations of Kirin chips, Huawei's cooperation with Cambrian is quite different.
At the end of 2017, at the beginning of the 970 Kirin Conference, the market had already caught a lot of clues about Huawei's layout of AI functions. Although the content of the subsequent conference mentioned that the new Kirin chip used AI core, Huawei not only refused to disclose its architecture source, but also sent congratulations to the Institute of Computing Technology of the Chinese Academy of Sciences on the formal mass production of the Cambrian Kirin chip. Huawei's attitude towards public congratulations is not readily accepted, but only emphasizes that its AI capability mainly comes from the accumulation of its own R&D for 10 years.
After the launch of Kirin 980 in September 2018, Huawei also tried to downplay the origin of its AI computing architecture, emphasizing only that it originated from Huawei technology.
What is more noteworthy is that Huawei's Qilin 700 Series in its middle end does not use AI architecture from Cambrian as in the 900 series, but uses ProjectTrillium from Arm. Huawei has cooperated with Arm to build the core of AI computing power and framework based on ARM computing architecture since 960. Although Qilin 970 changed to Cambrian architecture, its cooperation with Arm has never been interrupted.
In addition, Huawei is also actively promoting its AI computing layout. The lifting scheme is a completely self-made AI computing architecture, not only for cloud computing, but also for all terminal AI. This also means that the status of Cambrian is not irreplaceable for industry.
Therefore, the Cambrian period is in a rather awkward position. At present, the mainstream framework is the ready-made results of other people's development. With the promotion of Google, Microsoft and even Facebook, supporting the acceleration of multiple frameworks has become the essential function of all programmable AI chips/computing architectures. Opening is the selling point of AI chips, and the development of algorithms is also the guarantee of AI chips. The important direction of competitive advantage of our products, but Cambrian focused on the development of AI's general computing hardware platform, but there is not much in the algorithm. This makes the Cambrian into an awkward situation: the original Cambrian is to introduce to replace other AI computing schemes such as GPU, FPGA, but on the contrary, it has become an alternative part of itself.
Most of China's terminal chip customers have adopted the existing architecture, which makes it difficult to change the ecology.
Apart from Huawei, we can see various ASIC computing architectures including DSP, GPU, FPGA, visual recognition and speech recognition. For these operators, meeting their business goals is the key point, and the source of the architecture is not important. Considering that the existing structure is already quite good, there is no need to replace it. Otherwise, more manpower and technology R&D expenditure will be increased, and the effect will not necessarily be better.
How to highlight its position in the crowded market is a difficult problem that Cambrian encounters at present. After all, AI computing architecture is relatively simple and easy to develop compared with other architectures. Chip architecture is no longer the leading role. The algorithms and frameworks running in chips, even a set of development tools, are the key points. If there is no way to convince existing market players that their own development system is more advantageous, how easy is it to change the existing AI computing ecosystem?
In fact, there have been many cases in AI computing industry that regard NVIDIA as a hypothetical enemy, but none of these vendors has really replaced NVIDIA. After all, NVIDIA, as an important promoter of AI industry, relies not only on the simple efficiency of architecture, but also on the integrity of software ecology, development tools, and the complete educational heritage from the bottom to the top. It spent more than 10 years during this period, and laid out all kinds of industries. Finally, NVIDIA gained its present market position. In addition, NVIDIA started with the global market as its starting point. The goal is not just to stick to a specific market. After all, ecology pays attention to scale. If it can not be recognized by customers, it will be more difficult to change or even establish its own ecology.
Huawei's Ascending AI Computing Scheme, or the Cambrian Period, will face the greatest challenge
In the field of professional computing, Cambrian and Sino-Tech Dawning cooperate very closely. They use Cambrian architecture to build Phaneron server for reasoning application optimization, which is different from general cloud-based AI server. Phaneron's reasoning computing service for ultra-low latency includes related products applied in film security, network, broadcasting media entertainment, etc. Manufacturing and automation, financial, medical, logistics, transportation, retail and emerging applications, providing real-time AI services with lower latency for service providers.
In addition, Phaneron has comprehensive support in hardware neuron virtualization, Cambrian deep learning instruction set, Caffe, Mxnet, TensorFlow and so on. It can meet the needs of various AI computing environments with more flexibility. According to Cambrian, its performance is first-class in the industry.
On the other hand, the Cambrian also helped Sinopec Dawn to build an intelligent platform for rapid development and application - "SothisAI". SothisAI is characterized by heterogeneous integration, simplicity, flexibility and rapid deployment. Through AI unified heterogeneous programming model, AI computing power dynamic real-time allocation algorithm and AI basic performance function, it can quickly respond and release the computing power required by AI. Strategically, SothisAI will serve as a platform to support its general AI services and industry AI services.
At first glance, the Cambrian seems to have a comparable structure with NVIDIA and a solid partner, which seems stable. However, the truth is often cruel.
Huawei's recently unveiled AI program aims to target the market for cloud computing and terminal AI computing. Huawei calls it full-scene computing capability. The Da Vinci architecture will not only be a truly proprietary AI chip architecture for cloud applications, but also set the largest hypothetical enemy as NVIDIA as it did in the Precambrian.
In fact, the word "own" used to be quite ambiguous for Huawei. For example, in recent two generations of Qilin chips, Huawei has adopted IP from Cambrian, but in marketing, they all emphasize self-research. They do not want the media to link the Cambrian and Qilin AI computing units. The reason behind this is that Huawei is laying out its own AI computing system. Wuji is only an emergency plan for the transitional period.
At the Huawei Connection Conference held in Shanghai on October 10, AI chips, which were widely spread in the industry, finally showed their true colors. They not only cut thoroughly with the Cambrian, but also overlap highly with the Cambrian market.
Huawei released two AI computing chips in the venue. The first AI chip is a high-end rising 910. Its computing power can reach 256 TFOPS. It is the largest single-chip AI chip in the world. According to Xu Zhijun, chairman of Huawei's rotating board, the computing power of NVIDIA V100, 910, even though it is "the closest" to 910, is not only great, but also much more than double. The plan based on the chip is expected to be officially launched in the second quarter of 2019.
At the same time, Huawei's rising plus based on 910 technology will become the largest distributed training system in the world so far. Xu Zhijun introduced that by linking 1024 910 computers to build an AI computer cluster, it can provide super AI computing power up to 256 PFLOPS, thus achieving unprecedented speed to train your model, no matter how complex it is.
Huawei's second chip, however, is a 310 rise in mass production. This SoC belongs to the Rising Mini Series. With a maximum power consumption of only 8 watts, its integer precision can reach 16TFLOPS. At the same time, 310 also integrates 16 channels of high-definition video decoder. It is the most powerful AI chip for edge-oriented computing products. It can also be used for data center training and reasoning.
More Rising Series will be launched in the future, covering consumer terminals that may include mobile phones, which also means that the Cambrian will disappear completely from Huawei's product line. Huawei will not retail chips abroad, but build its own solutions and provide customers with the sales and services of the solutions.
Huawei's approach to server AI computing is the same as that of Cambrian. It not only promotes chips, but also actively promotes eco-management. Therefore, in the future, customers of Cambrian or Cambrian will launch solutions. Its biggest rival in the Chinese market is not NVIDIA, but Huawei.
Next step in the Cambrian?
As a supplier of core architecture, the most desirable thing is to increase the number of authorized customers and build a solid ecosystem. However, the Cambrian period has not achieved very well in this respect. Huawei is the only customer on the consumer side, while the server side has not yet been mass produced. But at this point, the Cambrian period also ushered in the possibility of losing customers, and determined to increase the double blow of the biggest opponent.
Cambrian as a new undertaking, only a few years of development requires a very large customer base is actually difficult to strong people. However, although the Cambrian concept of technology is good, it does not have a stable and reliable profit model to form an ecology, and the AI industry is changing with each passing day, which is very different from the original design of DianNao related architecture.
The Chinese market is eager for the Cambrian, hoping to bring different air to China's AI industry. But now different external computing architectures have long dominated the Chinese market. Even in the Chinese market alone, the choice of AI solutions is as many as the cow's hair. In addition, Huawei has turned into the biggest competitor in the past. The Cambrian wants to make a way out in such a dangerous market with no prospects. Optimistic.
Perhaps Cambrian should reconsider whether its IP licensing model is suitable for the characteristics of Chinese market and readjust its business strategy so as to take a relatively mature development ecology, re-enter the closed model, and try to win over the international big customers who want to do business in China, not only to create revenue, but also to create visibility, otherwise the existing business model will soon come to an end. (Sohu Technology)