Artificial intelligence helps chip design engineers to analyze the feasibility of placement and routing

Recently, a research team from Tsinghua University and other units published the design of using artificial intelligence based on reinforcement learning to automate millimeter wave circuits (the example in the paper is a filter). According to the results of the paper, the filter design implemented using artificial intelligence algorithms can reach indicators similar to those of real engineers. What is interesting is that the shape of the filters obtained by artificial intelligence algorithms is more irregular (a and d in the figure below are real engineers) The design, b and e are the design of artificial intelligence algorithm).

Recently, a research team from Tsinghua University and other units published the design of using artificial intelligence based on reinforcement learning to automate millimeter wave circuits (the example in the paper is a filter). According to the results of the paper, the filter design implemented using artificial intelligence algorithms can reach indicators similar to those of real engineers. What is interesting is that the shape of the filters obtained by artificial intelligence algorithms is more irregular (a and d in the figure below are real engineers) The design, b and e are the design of artificial intelligence algorithm).

Artificial intelligence helps chip design engineers to analyze the feasibility of placement and routing

Related reports have attracted attention in the field of circuit design. Many friends are surprised that the algorithm can automatically design filters with indicators that are close to those of real people, and hope that there will be more such artificial intelligence-based automated algorithms to accelerate circuit design in the future.

In my opinion, the use of such reinforcement learning artificial intelligence algorithms for filter design is indeed an academic breakthrough, and we can expect to see more such breakthroughs in the future. On the other hand, in fact, artificial intelligence has been highly valued in the EDA field, and even some of the tools we are currently using have already added artificial intelligence to the algorithm. However, we also believe that artificial intelligence is still a long way from truly automatically designing complex circuits above medium scale. In the future, artificial intelligence in the EDA field will mainly assist engineers in designing, simplifying processes, and improving design quality.

The nature of artificial intelligence

To understand the impact of artificial intelligence on EDA, we must first understand the nature of this wave of artificial intelligence algorithms. In the final analysis, this wave of artificial intelligence algorithms is a machine learning algorithm based on big data mining, which can grasp the statistical law among high-dimensional data from massive data, and realize more applications based on the statistical law.

In fact, what artificial intelligence is best at is the analysis and optimization tasks related to high-dimensional data. What is “high-dimensional data”? To explain it in a simpler way, it can be roughly understood as “variables affected by many factors.”

For example, if we take “the probability of a user buying a certain online red lipstick” as a variable, then we will find that the variable will be related to a large number of factors such as the user’s gender, age, education, income, place of residence, whether or not to like online shopping , To form a vector of all these related factors, it is a high-dimensional data. Humans are inherently difficult to quantitatively process this kind of high-dimensional data, so they often need to reduce the dimensionality. For example, according to the data, it is concluded that “satisfy women, age 20-25, college degree or above, monthly income above 5000, living in second-tier cities and above, like online shopping People with these conditions are most likely to buy this online celebrity lipstick”, but this is far from quantitative, because it is difficult for humans to find a fitted quantitative relationship between variables and influencing factors.

For example, the age of user A is 26 (beyond the aforementioned 20-25 range), but women who are far more enthusiastic about online shopping may not fall within the aforementioned regularity range, but they are even more likely than those who fall within the aforementioned regularity range. Most likely to buy this lipstick. However, if you use machine learning to find out the fitted quantitative relationship between variables and various influencing factors, for the example of user A, although his age item score is slightly lower, his online shopping enthusiasm score is extremely high, so he buys lipstick The total score of tendency will also be higher, and the system can smoothly classify it as potential consumers of the lipstick.

If we look at the previous example of artificial intelligence designing filters, we can see a similar situation. The routing of the filter is actually a high-dimensional data, because each pixel has a degree of freedom for the engineer to decide whether the routing should be covered here. For human engineers, because it is difficult to deal with such a high data dimension, the method of dimensionality reduction is adopted when designing, that is, to draw straight lines, and the adjustable variables are simplified to the thickness of the straight line, the distance between the straight lines, and the straight line. The number and so on. And artificial intelligence is not limited by the data dimension, so it will explore more free design patterns, and accordingly its design results will appear more irregular in humans-because once the rules are followed, it means that the data dimension will be reduced. .

Of course, although this generation of artificial intelligence has the powerful ability to process high-dimensional data, it also has strong limitations, that is, it lacks abstract reasoning, causal attribution and induction capabilities, and its generation ability is also weak. In addition, because artificial intelligence is data mining based on big data, it is difficult to work when the amount of data is not large enough. Therefore, artificial intelligence has not yet learned a set of abstract and universal design methods, which limits its ability to automatically complete complex designs.

Application of artificial intelligence in circuit design

According to the previous analysis, we can find that the most suitable application of artificial intelligence for circuit design is the exploration of high-dimensional data space, including problems that are difficult for human designers to grasp with massive influencing factors, and heuristic optimization problems.

The most common application example of machine learning and artificial intelligence in the EDA field is the placement and routing of the back-end of digital circuits. The placement and routing of digital circuits is a typical heuristic exploration problem, that is, because the variable space is too large, it is difficult to find the global optimal solution. Therefore, the actual method starts from a tentative initial solution and slowly explores and tries. Eventually it converges to a better solution. Friends who are familiar with the digital back-end placement and routing process will know that the usual routing process will include trial route, global route, detailed route and repeated incremental optimization. This is a typical process of finding a better solution through heuristic algorithms. .

For larger designs, each step of the routing algorithm has to run for a very long time (several hours to days), but in the end it may not really converge to a better solution, and sometimes the optimization results may even worsen. Phenomenon, this is actually a typical phenomenon of traditional heuristic algorithms, because the exploration of heuristic algorithms only has a high probability of getting better solutions, but there is no guarantee that a better solution will be found every time you explore.

So how can machine learning and artificial intelligence help place and route?

In fact, the help of machine learning here mainly comes from the ability to learn from big data. Traditional heuristic algorithms essentially do not consider the specific context of algorithm application, and theoretically explore various directions with equal probability each time in the process of exploration. Although in decades of engineering practice, people will add various optimizations to the wiring heuristic algorithm, but human engineers have limited grasp of the high-dimensional and complex problem of wiring. For example, when running the wiring, it is found that the local wiring is very crowded, and the wiring is loose at a certain distance on the right, so how many local wiring should be moved to the right?

If there is too little movement, it is difficult to completely alleviate the problem of wiring congestion, and too much movement is nothing more than moving the local congestion to the right, which is not optimal. This kind of problem of excessively high variable dimensions is difficult for humans to grasp, and machine learning can learn how to optimize the exploration process of heuristic algorithms from a large number of actual cases of placement and routing, so as to make the optimization results of the routing process better. At the same time, it can also reduce the number of explorations, that is, reduce the running time of the algorithm. Cadence announced last year that it had added machine learning elements to the place and route engine. As a result, it was able to improve the total negative slack of the running results by 15%. This can be said to be a good start.

In the field of place and route, another important application of artificial intelligence is to add predictive capabilities to the process, thereby reducing the number of iterations. As we all know, digital circuit layout is an iterative process, and the links before and after the process will affect each other. The layout process will affect the result of the layout, and if the result of the layout is too bad, it will cause the engineer to optimize the layout again for improvement. Such repeated iterative processes will obviously affect the design time, so can you predict its impact on the wiring during layout, so that you don’t want to optimize the layout after the actual wiring results come out?

There have been many attempts before, and the use of machine learning to make predictions by learning a lot of the relationship between layout and wiring congestion can greatly improve the accuracy of predictions, thereby reducing the number of iterations. At ISSCC 2017, Cliff Hou, TSMC’s vice president of research and development, focused on this machine learning-based back-end design method in his speech, and believed that it will bring more efficiency improvements to future integrated circuit research and development.

Artificial intelligence helps chip design engineers to analyze the feasibility of placement and routing

Artificial Intelligence and Circuit Designer

In the previous analysis, we can see that the current main advantage of artificial intelligence lies in the processing of high-dimensional data and the ability of quantitative statistical analysis. This wave of artificial intelligence is still weak in abstract reasoning, so it is difficult to fully meet the complex design tasks. When we humans deal with complex design tasks, the common method is to reduce their dimensionality, break them into multiple simpler sub-tasks and divide and conquer. When all the sub-tasks are completed, the complex design task is naturally completed. However, artificial intelligence currently lacks this adaptive ability to divide and conquer complex tasks.

If this ability to abstract simple laws from complex objects is called “dimensionality reduction”, the ability to grasp the complex relationships between high-dimensional data from a complex world can be said to be “dimensionality”. Humans have strong dimensionality reduction capabilities, while artificial intelligence has strong dimensionality enhancement capabilities. Therefore, the best way for humans and artificial intelligence to cooperate is for humans to be responsible for dimensionality reduction, abstract the design framework, and decompose complex designs into simpler subdivisions. Task, and artificial intelligence fully explores the high-dimensional design space in each sub-task to help complete the optimal design.

In this sense, artificial intelligence will not replace real engineers in the short term, but will help improve efficiency in the design process and reduce unnecessary iterations at the algorithm and design process level. All this is for real designers. Said it is all good. Therefore, we believe that artificial intelligence will be a good friend of circuit engineers.