Zhi DongXi (public number: zhidxcom)
Compile the | Zhao Di
Edit | Yunpeng
According to the IEEE, with the application of AI in chip design, the chip performance improvement cycle is gradually shortening, which may break Moore's law that "chip performance can be doubled every 18 months".
Heather Gorr, senior product manager of MathWork's MATLAB platform, said that AI technology can help improve the efficiency of chip testing, which is significantly faster than manually troubleshooting anomalies. The surrogate model based on AI design can adjust the parameters in the experimental data, simulate the experimental results, and realize the rapid update iteration of the model.
But the proxy model is less accurate than the physics-based model, and it also requires someone to integrate its results. Even so, the use of AI in the chip design process can indeed greatly release labor costs, use manpower for more advanced tasks, and achieve mutual assistance between technology and people.
First, use AI to process massive data and quickly find the cause of the problem
Before using AI for chip design, engineers and designers could only improve the performance of chips by shrinking the transistors, and now, AI is gradually becoming the key to improving chip performance, such as Samsung's use of AI in its memory chips to achieve faster and more energy-efficient memory processing, and Google's TPU V4 AI chip processing speed is also twice as fast as the previous version.
Gower said that AI can be applied to most aspects of the chip product life cycle, especially chip design. Because even if the design process takes into account a variety of possible problems, there will still be many abnormal situations in product testing, if you rely on manual analysis of the possible causes of problems in massive data, then the workload will be larger and slower, then you can use AI to complete tasks such as exploring the frequency domain, data synchronization or resampling.
In this process, AI is often seen as a tool for predicting problems, but many times engineers can also take inspiration from the information provided by AI, find problems that have not been noticed before, and come up with new ways of logical modeling. Gower stressed that many people have published relevant program code on the GitHub and MATLAB platforms, and engineers can directly use these tools to complete related tasks and improve work efficiency.
Second, the agent model saves calculation time and improves the efficiency of product optimization
Gower introduced ai to the IEEE on how AI can be used in chip design.
Previous chip design work is based on physical experiments, but AI can provide a proxy model, the input data for parameter scanning, optimization, Monte Carlo simulation (random sampling simulation method), etc., compared to the actual circuit circuit adjustment and other work, the agent model in the optimization of performance time is much less time.
IEEE said that in a sense, the AI-based proxy model can be regarded as a digital twin product, that is, on the basis of the physical model and experimental data, the proxy model adjusts parameters and simulates experiments to achieve rapid iteration of the model.
However, Gower points out that AI-based proxy models tend to be less accurate than models based on physical methods, which is why proxy models are required to perform multiple simulations and parameter scans. Proxy models can predict problems that may arise at every detail of chip design and manufacturing, but humans are also required to integrate these results.
Third, AI can not cover the chip design, but also need human participation
Gower describes what engineers and designers need to do when designing chips with AI technology. For engineers and designers, they only need to determine the problem they want to solve, and then they can give the specific solution to the AI, let the agent model adjust and simulate the parameters of each component, and finally the engineer and designer record the optimal solution parameters.
This frees engineers and designers from a part of the tedious work and greatly frees up human capital. Goyle said that using AI for chip design can reduce the waste of resources and optimize chip design, but she stressed that in the final decision-making process can not be separated from human participation, AI-based agent model can help people complete a lot of work, but its utilization depends on how people use it to achieve technology and human interaction.
Conclusion: AI breaks into the chip industry, improving the level of intelligence and expanding the application field
At present, GitHub and MATLAB platforms have announced a number of programs related to chip design, and more and more companies in the industry have begun to incorporate AI into chip design work, freeing up labor costs and improving work efficiency. Ai-based proxy models have played an important role in speeding up model updates, and the time required for chip design has also been shortened, and in the future, people may not have to wait 18 months to see chips that double their performance.
With the development of science and technology, AI more and more penetrated into people's lives, from smart home, language processing to health management, the application of AI is everywhere, in the field of extremely high precision chips, AI can also get a piece of the pie, Google even developed a tool to let AI code independently, in the future, AI intelligence and application scope will reach what height and breadth, it is worth looking forward to.
Source: IEEE