From algorithm to calculation, capital circle hot transfer
In 2017, when the capital circle and the media were in full swing, the consumers were confused at various AI technical forums.
The author Baidu a bit 'AI', found a film directed by the great director Spielberg, the AI in the movie is a little boy with no difference between the internal structure and the people.
So the people suddenly realized that, oh, the original AI is a robot?
There is also a movie called Matrix in the Matrix. The AI in this movie is the brain of a huge virtual world. The entire physical world becomes data, and humans themselves have become a series of code.
The two films actually reveal two manifestations of AI: cloud and terminal AI.
In 2016, most of the capital circles were focused on companies that focused on cloud computing algorithms, such as Shang Tang, Despise, Yun Cong, and four algorithms companies. According to Li Kaifu, they called the 'four unicorns' for face recognition.
In 2016, Intel acquired Nervana, Altea, Google released 'TPU' at the I/O Conference, Qualcomm began developing neural network chips, and Nvidia's stock price soared due to the concept of AI.
In 2017, the capital circle began to pay attention to the horizon, Shen Jian Technology, Cambrian, Yun Tian Lifei four AI chip companies.
The sense of smell of capital is very sensitive, and AI as a technology has been in existence for decades. The three major elements of AI include algorithms, computational power, and data. The accumulation and evolution of algorithms and data has continued for many years. The popularity of terminal computing has only just begun. From the point of view of computational development, it is always the first time to rely on algorithms to solve problems because the computing power is limited at that time. So capital focuses on algorithm companies. Now that computational power has risen, algorithmic differences have been To make up for it, so capital began to focus on companies that provide power.
Although AlphaGo is incredibly powerful, it is based on the use of thousands of CPUs and hundreds of GPUs. The cost of electricity for the next game is as high as $3,000, which is quite energy-intensive.
The mobile phone is the most popular personal mobile terminal. How can a low-power smart phone be as smart as AlphaGo?
Why does Huawei Kirin lead in the introduction of mobile AI chips?
The mobile phone industry is the first to act on two companies - Apple and Huawei. In the iPhone 11's A11, the Neural Engine was used as a neural engine. After years of technical accumulation and advance research, Huawei was innovative. The first to carry the NPU (Neural-network Porchinging Unit) in the Kirin 970.
Huawei Kirin Chip Marketing Director Zhou Chen
'We are fortunate enough to get to the front of this matter (AI). Our Kirin 970 and iPhone X's A11 are the first companies to truly introduce mass-production products on the terminal side. 'Wealthy Chip's Marketing Director Zhou Chen told EE Times' "Reporter said. Compared to Apple's AI chip, Huawei Kirin is using hardware to speed up, 'We have something different for the open, Apple's ability to use this mainly in FaceID, we are fully open to developers through HiAI Developers can use the industry's strongest AI reasoning capabilities. ' Zhou Chen said.
As the benchmark of the mobile phone industry, the reason why these two companies can take the lead in introducing AI mobile phone chips is, of course, because they each have chip design capabilities. The second reason is that their chip companies or companies only need to serve one customer.
Actually, the AI project was started 2 years ago. Because we are a chip company in house, we only need to serve Huawei brand. We can clearly know what the terminal needs are, so that we can make products targeted. And planning and designing. Zhou Chou thinks that the logic for the open market chip companies is not the same. Because adding any function in the chip is a cost, pure chip companies need to consider that their customers' demand for AI will be How, therefore, it will be conservative. In the future in the adoption of new technologies, similar to Apple, Huawei in house will be more radical, this may be Huawei's first in the industry to launch AI mobile phone chip an important reason.
From cloud to end, how does an AI chip iteration satisfy the algorithm change?
Of course, if you think that Apple and Huawei are using the 'AI' module in the mobile phone chip just to find new selling points, as a marketing gimmick to sell more mobile phones, then you're in the picture. From the cloud to the terminal can really bring real benefits. , Such as the reduction of power consumption and cost, the improvement of efficiency and so on.
The 2018 CES show confirmed the trend of AI transfer to the terminal. Many of the previous concept-based applications began to really take the form of products and services. From the perspective of AI products exhibited at CES, there are roughly three categories: One type is the autopilot program applied to automobiles. At present, Nvidia is mainly deployed in this field. The second category is related to audio-based assistant services, including smart speakers, etc. The third category is the product of image recognition. Startup companies are entering, developing new applications like security monitoring, live streaming, video effects and more.
Zhou Chen believes that the way to perform AI operations in the cloud, such as increasing the size and number of layers of the model, is not the real direction. AI began to migrate from the cloud to the terminal, which will reduce the amount of operations and network size by 100 times, and will also reduce the network bandwidth. Demand. 'With the improvement of computational power and optimization of the algorithm, a good sweet spot will eventually be formed.' Zhou Chou said that after the CES period this year, he went to the Silicon Valley to meet and found that the industry was discussing On device. Started AI. There are many companies in the company, and there are many chip companies.
'The performance of the terminal is now a bottleneck. Although a large number of companies have uneven levels, they are all moving toward the chip. How to solve the compatibility problem and how to make the algorithm smoother have become the focus of everyone's attention. ' Zhou Chen stated that from 2017 According to the data, AI's computing power accounted for 95% of the cloud, and the terminal accounted for only 5%. The huge gap between the two points will bring a lot of room for growth.
Of course, due to the current AI algorithm update speed is very fast, how to meet certain variability? Zhou Chen believes that the future more appropriate architecture should be CPU + hardware accelerator approach. For Huawei Unicorn, because it is a hardware platform company, naturally It prefers to use hardware to provide universal accelerators to meet most of the algorithm requirements. Currently there are two ways to run the algorithm: one is to run pure software with the CPU, and the other is to use hardware accelerators to fix Algorithm. In addition, from the perspective of the requirements of AI operations, most of them require high computational density, and they also need to be done in a very efficient hardware manner. Zhou Chen believes that from the history of chip development, many algorithms once Stable and mature will be fixed with hardware.
'We have our own GPU, DSP resources, and after doing a good job with heterogeneous things, the unknown application for third parties is accelerated as much as possible. Because the mobile phone is an open platform, so we must do this. ' Zhou Chen said, Qi Lin The positioning is the provider of platform computing power. At the same time, SDK will be provided for developers to run their own applications and algorithms. 'Upwards may Huawei provide some business-level interfaces on the mobile terminal, these interfaces do not require developers to come To do algorithms, such as calorie identification, APP simply calls the interface directly.
What about the next generation of AI chips?
For Huawei’s next-generation AI chips, Zhou did not disclose too specific specifications or parameters. However, he said that the definition of the products of the latter two generations is clear and basically there are several directions that will not change: The first is that the computing power will continue to rise. The second is to focus on the common AI computing platform, and continue to open up AI computing capabilities to more developers. Third, the process of AI chips will be more and more advanced, and the update speed may be faster than Moore's Law.
'When we release the computing power, there may be a big boost in application capabilities. This will create a positive loop. There will be more and more developers doing some good new experiences.' Zhou Chou said, AI itself Is a technology, but the future value will be reflected in how many developers to develop related applications.
At present, there is a point of view in the industry that AI deep learning requires data to feed, so Internet companies with larger data volumes, such as BAT, will occupy core advantages in the future AI field. For this statement, Zhou Chen believes that the unicorn 970 computing power and algorithm In fact, all have advantages, as far as the data is not a problem.
The first reason is that the algorithm itself is upgrading, and the efficiency of machine learning itself is improving. 'The way of deep learning differs from the way the brain learns. For example, when a child watches a few cats, he can know what a cat is. You don't need to look at 100,000 photos to learn. Zhou Chou thinks that the human brain is actually doing a higher level algorithm and extracting higher-dimensional features. Therefore, when the algorithm is upgraded, it will no longer need so much data. Do deep learning.
The second reason is that the terminal has become a human computing center. The data generated on the mobile phone is truly one-handed, fresh data, and the data in the cloud may be compressed or streamlined second-hand data.
Zhou Chen believes that after 2018, AI will become the standard demand of major mobile phone companies. The huge amount of data on the mobile phone terminal will be used. Only after the terminal AI hardware capabilities are popular, can AI applications flourish and the AI ecology can flourish. It may be really mature.
What is Huawei 'HiAI'?
For chip companies, how to meet the different application requirements of the AI and the differentiation needs of the ecosystem has become a difficult and important point. 'Huawei will also provide a HiAI architecture, and we will also have an open class of HiAI held on a regular basis, ' Zhou Chen last Huawei said that Huawei has held an open and open course for Huawei HiAI capabilities for engineers. It will help companies and developers trying to join the field of artificial intelligence, increase the competitiveness of technical personnel, promote the promotion of industry technology, and jointly promote the development of artificial intelligence ecosystem.
What is Huawei HiAI? It is understood that the HiAI mobile computing platform has three layers of capabilities, namely, Huawei HiAI Services, Huawei HiAI Engine, and open-end chip capabilities (Huawei HiAI Accelerator). ).
Among them, the underlying Huawei HiAI Accelerator is the core content of the HiAI chip's capability opening. It can quickly convert and migrate existing models. With NPU's acceleration, it can achieve the best performance. It mainly relies on Unicorn chips. The goal of Huawei's HiAI engine is through the open application layer API. Enabling developers to develop high-quality AI applications without knowing AI algorithms, focusing entirely on the application experience and business practices, rather than focusing on a large number of back-end model training and algorithms.
In the fiery AI field, the rapid introduction of high-quality AI applications is the best way to seize market opportunities. This time, Huawei HiAI Engine is able to cooperate by opening and sharing Huawei's own application layer commercial-grade AI capability API. Partners will win the competition in the future competition, enabling applications on Huawei mobile phones to provide the industry's best AI experience. Therefore, Huawei HiAI Engine provides multiple AI capabilities and JAVA application layer interfaces for mobile terminals.