Artificial intelligence (AI) set off a new wave of technology, making the market crazy. UBS (UBS) predicts that four chipmakers, including Nvidia, AMD, Intel Corp and Qualcomm Inc, will dominate the entire AI industry in the future, of which Nvidia is the most promising.
Stephen Chin, an analyst at UBS, pointed out that the development of AI is just in its infancy, and there is still a lot of room for growth in the future, which will push the semiconductor industry into a new chapter. The UBS model predicts that the output value of semiconductors (excluding memory) related to machine learning and artificial intelligence will grow from $8.2 billion this year to $35 billion in 2021, a compound growth rate of 41 per cent. (Businessinsider)
UBS analysts are most optimistic about Nvidia because Nvidia technology is ahead of the industry for at least a year, and Intel Corp is the only chip factory with a full range of AI solutions. Ultramicro, Qualcomm Inc, Samsung and even Google technology are not enough to cover the entire AI industry.
However, some analysts hold different views. Morningstar analyst Abhinav Davuluri pointed out in a report released at the end of September that although Nvidia entered the AI field early, it does not mean that Nvidia can dominate forever and may still be overtaken by latecomers in the future.
NVIDIA Corp, the biggest winner in the era of artificial Intelligence
NVIDIA Corp, who failed on the mobile a few years ago, should not have thought that he would be a big hit in the artificial intelligence market today. With the advantage of GPU, they will undoubtedly become the big winners of artificial intelligence.
According to Caixin, from the perspective of computer development, the original birth of GPU is related to the growing video game market, and it is obviously different from CPU. The former is used to solve tasks with complex computing steps, while the design of GPU is more suitable for batch operation of massive data. The latter is exactly what the deep learning technology needs at the core of the artificial intelligence era.
After Moore's Law slows down, GPU accelerated computing expands computing power, allowing applications that desperately need more computing to continue. "Huang Renxun said. Moore's Law states that the number of transistors that can be held on an integrated circuit doubles about every two years. The industry believes that Moore's Law will last until around 2020. "nowadays, the scope of application of GPU is getting wider and wider, and it has been involved in various fields. "
Huang Renxun described that AI is "eating" software, and all aspects of software in the future will be infiltrated by AI. And NVIDIA Corp has long been aware of such opportunities from the perspective of upstream suppliers and laid out in advance.
Such a change is also the expectation of the market for NVIDIA Corp. Artificial intelligence is the most popular concept on Wall Street today, and what excites the market even more is that at a time when major technology companies are losing money, NVIDIA Corp has begun to make money. In the first quarter of this year, NVIDIA Corp's data center business revenue of US $409 million was much better than the expected US $318.2 million, an increase of 38% from the previous quarter, and achieved growth for the seventh consecutive quarter.
"all the major Internet and cloud service providers in the world are using NVIDIA Corp's Tesla GPU accelerator," NVIDIA Corp CFO Collette Kress said on the earnings call. Artificial intelligence is rapidly becoming the most influential force in the field of science and technology, and it is NVIDIA Corp's GPU who stands at the center of artificial intelligence. "
Cars are another focus of Nvidia.
Since last year, NVIDIA Corp has successively established cooperative relations with Mercedes-Benz, Audi, Bosch, Tesla, Inc. and Toyota. NVIDIA Corp also successively released the next generation of Xavier in-car computers, BB8 driverless cars using the DRIVEPX2 on-board computer platform, and AICo-Pilot, an artificial intelligence cooperative driving system with multiple perceptual functions.
NVIDIA Corp's first collaboration with some automakers began with in-car entertainment systems, and now it provides a "brain" for cars, and the change over the past decade looks like a counterattack. NVIDIA Corp is undoubtedly the darling of this artificial intelligence era, and it can be found in all the hottest tuyere. It is said that the chip industry in the PC era belongs to Intel Corp, the mobile phone era belongs to Qualcomm Inc, is entering a new era of artificial intelligence, and NVIDIA Corp has taken the lead.
Today NVIDIA Corp is a GPU computing company engaged in artificial intelligence business, but one day we hope to directly become an artificial intelligence computing company. Huang Renxun said in an interview with Techcrunch.
AMD: always the second dream of counterattack
Both CPU and GPU,AMD are second, and they have been planning a counterattack all the time. In the age of artificial intelligence, it seems that their dream of counterattack has been dashed again.
AMD recently announced new measures to achieve these goals. The first is a new graphics card product, Radeon Vega (based on the previously announced new graphics card architecture), and the second is the new open source software platform ROCm-- that enables machine learning frameworks and other applications to take advantage of the software layer of multiple graphics cards.
These two elements, hardware and software, are equally important. For AMD to fight back against NVIDIA Corp's advantages in the field of machine learning, both are indispensable.
AMD new generation star graphics card: Vega
AMD has long been committed to providing the most cost-effective products, whether processors or graphics cards (or the long-rumored two-in-one product). The goal of Vega--AMD 's new generation of graphics cards is not to become a cost-effective substitute for NVIDIA Corp Pascal series graphics cards, but to completely defeat Pascal.
InfoWorld said that the early running scores disclosed by AMD showed that the score of the Radeon Vega Frontier Edition card (a professional Vega card) in DeepBench was 1.38-1.51 times higher than that of NVIDIA Corp's Tesla P100 card-related to the version of NVIDIA Corp driver used.
Although running points doesn't have to be taken too seriously, such a large performance gap is still impressive. Also important is the price of AMD products. The Tesla P100 retails for about $13000 (89000 yuan), and AMD has not disclosed the Vega Frontier price. Even if the price of Vega Frontier is comparable to that of the Tesla P100, it is still very attractive and in line with AMD's overall business strategy.
AMD's Technology for dealing with CUDA: ROCm
What is more important for AMD to gain an advantage in machine learning is not to beat NVIDIA Corp in price, but to ensure that its hardware gets at least the same level of support as NVIDIA Corp in common machine learning applications.
In general, the software that uses graphics cards for acceleration uses NVIDIA Corp's CUDA library files-- only NVIDIA Corp hardware is supported. The open source OpenCL library provides vendor-independent support on many types of devices, but its performance is not as good as specialized solutions such as CUDA.
Instead of trying to improve OpenCL to rival CUDA, a slow process driven by the committee, AMD created its own open source graphics computing platform, ROCm (Radeon Open Computing platform). The idea of AMD is that ROCm provides a middleware layer independent of language and hardware for graphics cards, mainly AMD's own graphics cards, which is theoretically suitable for any graphics card. ROCm can also communicate with the graphics card through OpenCL if necessary, but it also provides a channel to communicate directly with the underlying hardware.
There is no doubt that ROCm can greatly improve the performance of machine learning applications compared with OpenCL, InfoWorld said. Porting the Caffe framework to ROCm is about 80% faster than the OpenCL version. In addition, AMD claims that porting code to take advantage of ROCm is a highly automated process, which is another "incentive" for existing frameworks to try ROCm. Support for other frameworks, such as TensorFlow and MxNet, is also being planned.
AMD's ultimate goal is not complicated: to create an environment in which its graphics cards can replace NVIDIA Corp products in machine learning. AMD can achieve its goal by providing hardware with equivalent or higher performance at a considerable price to ensure that the existing machine learning software ecological chain can run on its graphics card.
To some extent, porting software is the easiest part. Porting software is basically about hiring enough programmers to rewrite the code needed for the most important open source machine learning framework, and then updating the code as the hardware and framework evolves.
Perhaps the most difficult task for AMD is to gain a foothold in large-scale applications that provide graphics cards. The graphics cards in Amazon Web Services, Azure and Google Cloud Platform are all NVIDIA Corp products. Other graphics cards are not supported yet on demand. However, if the new generation of machine learning software is more independent of graphics cards, cloud service providers will lose an excuse not to use Vega or its successor products.
Any plan by AMD to guide the needs of its graphics cards in the field of machine learning is bold. It will take years for AMD to succeed, because it is facing a world that NVIDIA Corp has dominated for many years.
Intel: semiconductor boss awakens
Like NVIDIA Corp, Intel Corp has struggled in the quagmire of the mobile market for a long time, but the emergence of artificial intelligence has made them full of vitality again.
As a company specializing in data processing, Intel Corp is also deeply aware of the coming great changes in the future AI field and the new requirements for data computing.
Intel Corp's judgment is that the hardware in the field of artificial intelligence will be more diversified in the future, but with the development of the computer age becoming more mature, the deployment of many technologies becomes very difficult, because many technologies are under the framework of artificial intelligence, but in the whole AI-related fields, only 7% of the applications meet the specific requirements and demands of AI.
Therefore, in order to better achieve artificial intelligence, Intel Corp is also constantly expanding his technology layout, including the acquisition of Mobileye, the world's leading provider of self-driving solutions, Nervana, a leader in deep learning and neural network chips and software, Movidius, a leading computer vision company, and Saffron, a leading artificial intelligence service provider. By combining these investments with Intel Corp Xeon, Xeon nuclear products, realistic technology and FPGA, it provides full stack power to process end-to-end data, from hardware, libraries and languages, frameworks, tools to applications, with all the assets needed to provide end-to-end artificial intelligence solutions to the market.
At the software level, Intel Corp said: "there are still many technical breakthroughs in the current deep neural network. Neural network is only a tool to achieve artificial intelligence, but it may not be the last tool." "
For the AI algorithm, Intel Corp is currently working on how to make the system rely on less data and manpower; how to make the model more sparse; how to compress the model to make it easier to store; how to cut the model to minimize the amount of computation; and how to reduce the calculation accuracy to one place. They believe that the algorithm of deep learning can be divided into two parts, one is changeable and the other is relatively reliable. The basic operation, task calculation and convolution calculation of deep learning algorithm have not changed much in these years, so the following deep learning technology will not change completely overnight.
Qualcomm Inc: an extension of the Mobile Chip Giant
With the influence of mobile chips and technology, Qualcomm Inc is very popular in the smartphone era, and they don't want to miss the artificial intelligence era. In addition to adding AI to his Snapdragon chip, Qualcomm Inc also develops artificial intelligence in other ways.
Not long ago, Qualcomm Inc announced the acquisition of Scyfer, a Dutch start-up affiliated to the University of Amsterdam that focuses on machine learning and deep learning technologies and solutions in the manufacturing, healthcare and financial sectors. Shen Jin, global vice president of Qualcomm Inc Group and general manager of the Investment Department in China, revealed that Qualcomm Inc's related companies and projects in the field of artificial intelligence include: American brain Group (brain), which took seven years to start from studying human brains to providing software for robots to achieve mobility and navigation; as well as American face recognition company Clarifai; China Intelligent Voice Company Yunzhisheng. But in fact, Qualcomm Inc's exploration in the field of machine learning began as early as ten years ago. At first, Qualcomm Inc mainly studied the impulsive neural method of machine learning for computer vision and motion control applications. After a period of time, Qualcomm Inc's research gradually expanded to the field of deep learning, including perception, reasoning and action, and covered listening, seeing, monitoring, observation, learning, natural interaction and privacy protection.
When we talk about artificial intelligence, there are no more than five elements: algorithm, computing, speech, image recognition and vision. Internet giants who can be called up at home and abroad are also stepping up the application of "cloud" in several of these areas. Unlike them, Qualcomm Inc does not focus on the AI in the cloud, but hopes to apply machine learning to the terminal, that is, to focus on his own board.
The role of many enterprises is crucial to the direction of artificial intelligence in the future, but the promotion brought by these leaders is undoubtedly worth looking forward to.