Hello, Community members.
Since Artificial Intelligence has gradually entered people's lives, AI chips have been developed in recent years.
After Huawei released the architecture of the Da Vinci AI chip, I have been puzzled about the AI chip. I have written down my understanding of the AI chip by reading some professional articles, hoping to help you.
This series mainly includes:
Talking from the Alpha Go win the Chinese Weiqi Association
ASIC Integrated Circuit Model Redefines the AI Chips
AI chips depend on Architecture Innovation
Huawei Da Vinci AI Chip Architecture
Preface
According to Huawei Global Industry Vision (GIV), the global data volume will increase by 180 ZB in 2025, far exceeding the human processing capability. 95% of the data will be processed by AI.
Data is an important asset of enterprises. Using AI to analyze, process, and make decisions more efficiently and improve production efficiency and intelligence will become one of the core tasks of enterprises. It is estimated that by 2025, the adoption rate of AI by global enterprises will reach 86%. The rise of AI will profoundly change the business model and value creation model of enterprises.
Over the past 60 years, AI has witnessed several changes but has been constantly making new breakthroughs driven by emerging ICT technologies. In recent years, however, the CPU performance has not doubled as predicted by Moore's Law. The general opinion in the industry is that Moore's Law has failed.
So, whether a chip with ultra-high computing capability and meeting market requirements can be developed has become an important factor for the sustainable development of the AI field.
Talking from the Alpha Go win the Chinese Weiqi Association
In 2016, Alpha Go and Go world champion lee sedol staged the "Century Man-Machine Wars", which pushed the attention of AI to an unprecedented level. The AI robot AlphaGo finally defeated lee sedol by 4:1. In this man-machine battle, Google DeepMind consumed 1,202 CPUs and 176 GPUs. The floating-point computing capability of AlphaGo is 30,000 times that of IBM DeepBlue when it defeated the chess champion in 1998. But from the perspective of energy efficiency, has Alpha Go really defeated mankind? We will analyze the following aspects. Adult males need about 2,550 kilocalories a day, 1 kilocalorie (KCAL) = 4.184 kilojoules (KJ).
If we convert calories into joules, it's about 10 million joules, and we play chess for an hour, li lee sedol consumes about 0.7 megajoules. When Alpha Go and lee sedol play chess, they use 1202 CPUs and 176 GPUs. If one CPU (100 W) and one GPU (200 W) are used in one hour, AlphaGo is required. One watt-hour equals 3600 joules. Therefore, a total of 559 megajoules are consumed, this is equivalent to the energy consumption of lee sedol, which is about 1/8 of that of Alpha Go.
Therefore, GPUs provide higher performance and efficiency than CPUs but are more suitable for large-scale distributed training scenarios. With the development of 5G, IoT, cloud, and ultra-broadband information technologies, intelligence will be extended to every smart device and terminal, including various forms of edge computing, IoT, and consumer intelligent terminals. To achieve optimal user experience, such devices are often deployed close to users and require long-time standby time, the power consumption and space requirements are high. Obviously, the GPU cannot meet the requirements of such scenarios. The essence of AI is to help all industries improve production efficiency and generate social and commercial value.
If Alpha Go relies on huge and expensive computing resources to implement a simple scenario, it is just a waste of resources. According to our understanding of AI requirements, AI chip R&D must cover all intelligent requirements in all scenarios, including cloud, edge, and device. Whether it's deep learning training, or reasoning, or both, it's not a chip package. From the perspective of AI chip development, it is also gradually adapt to this process.
That's all for today. I hope it will be helpful to all of you!
If you have any other comments or want to know more, please leave a message below to let me know.
