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

Introduction: technological innovation is closely related to the computing power that carries it. Today's artificial intelligence technology is gradually moving from research and experiment to application and production. In this process, the importance of AI computing system design and optimization is becoming more and more obvious.

The innovation of technology is closely related to the computing power that carries it. Today's artificial intelligence technology is gradually moving from research and experiment to application and production. In this process, the importance of AI computing system design and optimization is becoming more and more obvious.

In the near future, AI computing system will face some challenges, such as optimal design of computing platform, computing efficiency in complex heterogeneous environment, high parallelism and expansion of computing framework, computing performance of AI applications, and so on. The challenges caused by the development of computing power to the whole computing requirements will become greater, so it is urgent to improve the performance and efficiency of the whole AI computing system.

Computing power is the pain point of artificial intelligence

The current artificial intelligence, more on behalf of intelligent individuals, through their own continuous learning ability, intelligent completion of single-point decision-making.

Machine experience needs to be derived from a large amount of historical data, so data collection is everywhere, and the growth of data will be geometric series. Using the current centralized storage and centralized communication mode, it is impossible to support such a large volume through a giant single point in the future, storage and communication capacity is the bottleneck, and the efficiency will be very inefficient.

Not only that, the cost of computing is also a major pain point for the artificial intelligence industry. Nowadays, the hardware investment of artificial intelligence enterprises is very large. Artificial intelligence has a great demand for computing, so there is a high demand for customized deep learning chips for high-performance computing, which means that many enterprises have to spend a lot of money to buy computing power and build a lot of computing centers, resulting in a great waste of resources.

AI computing power is becoming more and more important

The business value of the global artificial intelligence market has been growing rapidly. By 2018, the business value will reach 1.3 trillion US dollars, and in the future it will reach nearly 5 trillion US dollars. From the technology maturity curve, we can see that in the next 2 to 5 years, a large number of AI technologies will achieve the transition from the innovation period to the growth stage, and there are still many AI technologies in the climbing stage.

In the process of promoting the development of AI, there are three major elements that play a leading role. In addition to computing power and data, computing is becoming more and more important. Computing investment will account for nearly half of all AI investment in 2021 and will grow nearly sixfold between 2017 and 2022.

Traditional servers are difficult to meet the computing power requirements of AI.

Moore's Law fails, and the performance improvement of CPU encounters a bottleneck. Intel announced the official suspension of the "Tick-Tock" processor development model, and the future research and development cycle will shift from a two-year cycle to a three-year cycle. The improvement of the performance of a single CPU is slowing down, the traditional server is difficult to meet the demand of parallel computing, and the growth of server CPU shipments is stagnant.

The Design of AI Computing system is extremely urgent

In the evolution of the development trend of AI computing, there are great challenges: with the higher the accuracy of the model, the amount of computation required will increase. For the future, the challenges caused by the development of algorithms to the whole computing requirements will become greater, so it is particularly important to improve the performance and efficiency of the whole AI computing system.

Around the characteristics of the whole application, computing power, model and network, for example, some model parameters are dense, the communication requirements of the system are high, and some computing performance requirements are high. We need to consider how to improve the performance of the whole system around the performance.

The complexity of reasoning is higher than that of training. it is not only reflected in performance, but also more concerned about the user experience. When deploying cloud computing on a large scale, we need to consider its operation and maintenance costs and need low-power platform architecture to support it.

The computing power and demand of AI are growing day by day.

Speech recognition, virtual reality and machine vision have entered the growth stage from the introduction period. Standardized data sets are rich. Voice and image data are easily tagged. In the past 15 years, the accuracy of image recognition has exceeded that of human beings. The basic computing power of CNN, RNN and other neural networks is mature. According to the Imagenet test results, the accuracy of speech recognition and machine recognition is more than 90%.

Nowadays, AI is facing huge computing challenges, so it is particularly important to improve the performance and efficiency of AI computing system, which needs to be considered comprehensively from a system point of view.

Since 2012, the amount of computation used in large AI operations has increased exponentially, and the amount of computation has doubled in 3.5 months. Since 2012, this indicator has increased 300000 times. The improvement of computing power is an important factor in the progress of AI, so as long as this trend continues, it is worth working hard for the future development of AI.

AI computing power determines the rise of chips

At the theoretical level, artificial intelligence technology based on big data and deep learning highly depends on the data processing and learning ability of the system. Therefore, the computing power of hardware has become another major bottleneck restricting the development of artificial intelligence after data and computing power. Hardware computing power and energy efficiency mainly affect the application of artificial intelligence technology in cloud and edge.

The development of single CPU can no longer meet the needs of practical applications, and the era of AI must rely on parallel computing. At present, the mainstream architecture of parallel computing is heterogeneous parallel computing platform.

Take the collaborative filtering computing power as an example, it is a very important recommendation system computing power, so two computing cores are designed, one is to calculate the similarity, the other is to calculate the average, all these core calculations are carried out on the GPU to speed up the computing power.

The Development of AI Computing Power in the Future

The amount of computing power used in artificial intelligence training tasks has been growing exponentially since 2012 and is now doubling every 3.5 months.

Since 2012, the demand for computing power has increased by more than 300000 times. At the rate of Moore's Law, it will only grow 12 times. During this period, the improvement of hardware computing power has been an important factor in the rapid development of artificial intelligence. Therefore, if the current trend continues, you need to be prepared to implement a completely new system that far exceeds the load of the current method.

From the perspective of application scenarios, AI computing popular industry scenarios cover the Internet, government, medical and financial industries. at the same time, according to the market potential and time maturity, the report also evaluates the typical application scenarios of AI, and predicts that in the next 2-3 years, the application of artificial intelligence in biometrics and smart city construction will be the first to enter the mature period of commercial applications. It is expected that in the next 5-10 years, the application of artificial intelligence industry in smart home and industrial manufacturing will gradually step into the industrial window of rapid development.

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.