The chip industry is undergoing qualitative change, and the boundaries between hardware and software are being broken by AI.
”
Author | Bao Yonggang
Edit | Wang Yafeng
Over the past few years, there have been some interesting changes in the market competition in the chip industry. In the PC processor market, Intel, the long-standing hegemon, is facing a fierce offensive by AMD. In the mobile phone processor market, Qualcomm has given up the throne of the first shipment for five consecutive quarters, and MediaTek is full of spirit.
When the competition of traditional chip giants intensified, technology giants who were good at software and algorithms began to develop their own chips, making the competition in the chip industry more interesting.
Behind these changes, on the one hand, is the slowdown of Moore's Law after 2005, and more importantly, the need for differentiation brought about by the rapid development of digitalization.
The performance of the general chip provided by the chip giants is reliable, and the increasingly large and diverse application needs of autonomous driving, high-performance computing, AI, etc., in addition to the performance, are more in the pursuit of differentiated functions, and the technology giants have to start developing their own chips to consolidate their ability to grasp the terminal market.
While the chip market competition pattern changes, it can be seen that the chip industry will usher in greater changes, and the factors driving all this change are the very hot AI in recent years.
Some industry experts say that AI technology will bring disruptive changes to the entire chip industry. Wang Bingda, chief innovation officer of Synopsys, head of the AI lab and vice president of global strategic project management, told Leifeng, "If it is to design chips with EDA (Electronic Design Automation) tools that introduce AI technology, I agree with this statement. ”
If AI is applied to a single link of chip design, the accumulation of experienced engineers can be integrated into EDA tools, which greatly reduces the threshold of chip design. If AI is applied to the entire process of chip design, the existing experience can also be used to optimize the design process, significantly shorten the chip design cycle, improve chip performance, and reduce design costs.
1
Chip industry "qualitative change"
Moore's Law has been in effect for more than 20 years, and chip companies have been able to achieve continuous and substantial improvements in performance and energy efficiency with the continuous scaling of transistors. Therefore, in the past few decades, hardware and software can be said to be "well water does not violate the river water", general hardware has a fixed architecture, computing power continues to increase, and products are updated in annual cycles. On the basis of the general chip, the system company innovates at the software level, and the product is iterated in weeks or even days.
"The trend now is that software and hardware are jointly designed, and the boundaries between software and hardware of chips are no longer so clear." Wang Bingda pointed out, "Breaking this boundary is the emergence of AI chips, because the architecture of AI chips is not as fixed as CPUs and GPUs, and the designers of AI chips can design special architectures and chips according to the needs of the application and combine general AI operators." ”
Sassine Ghazi, global president and chief operating officer of Synopsys, also said that under the digital trend, large system-level companies have developed their own chips and customized chips to optimize their applications or workloads. In the Chinese market, segments including the electrification and driverlessness of automobiles, AI, and ultra-large-scale data centers are undergoing major transformations, and they all want to differentiate their systems by customizing SoCs to find the differences in the overall business and have a differentiated competitive advantage. Only with better chips can they differentiate their system architecture. Domain Specific Architectures (DSAs) reflect the unique strengths of their system architectures.
That is to say, the domain-specific architecture allows chip designers to decide that part of the algorithm becomes hardware, and part of the algorithm continues to use software, in a more flexible way, through better collaboration between software and hardware, and more efficiently meet the needs of the final application. In this way, architectural innovation has become the key to the next competition in the chip field.
In early 2019, two Turing Laureates, John L. Hennessy and David A. Patterson, published a lengthy report, A New Golden Age for Computer Architecture, in which they looked ahead to the next decade as a "new golden decade" in computer architecture. In Wang's view, this may require new EDA tools that can automate architectural exploration, such as integrated deep learning accelerators, to better adapt to the needs of specific applications.
"The change of architecture will bring a lot of uncertainty, and the architecture of the general chip in the past is determined, mainly in terms of process improvement." Wang Bingda said, "The SysMoore concept proposed by Xinsi should take into account all factors from architecture to process to system level, and the changes and uncertainties brought about by it cannot be solved by traditional methods, and AI can play a great role." ”
In addition to uncertainty, architectural innovation also requires that the cycle time from design to production of chips should be significantly shortened, otherwise it is difficult to quickly meet changes in demand.
As early as 2018, the U.S. Defense Advanced Research Projects Agency (DARPA) proposed two new projects, IDEA (Intelligent Design of Electronic Assets) and POSH (Posh Open Source Hardware), with the goal of shortening the chip design process from the two dimensions of IP and EDA and saving R&D time.
The convergence of AI and EDA can fundamentally solve these challenges.
2
AI technology will disrupt chip design
In June, the Google team published a paper titled "A Chip Design with Deep Reinforcement Learning" in nature, a top international journal, which pointed out that with deep learning, human engineers need to complete the work in months, and Google can achieve the same effect in just 6 hours with AI, an improvement of hundreds of times.
Wang Bingda said: "Using EDA tools with AI technology to design chips, the time will definitely be shortened, this is beyond doubt, but the degree of time reduction is different." ”
The reason why AI can shorten the chip design cycle is not complicated, mainly to let AI first learn, with the accumulation of knowledge, in the subsequent use of the same or similar problems can be solved at a faster speed, so EDA with AI can save chip design cycle is almost a foregone conclusion.
Ai is applied to EDA in two forms, because chip design is a long and complex process, the whole process may require more than a dozen EDA tools, so AI can be applied to the EDA point tool to optimize a single chip design link, can also be used for the optimization of the entire chip design process.
If used in a single EDA point tool, it plays a role in sharing experience, enabling an engineer with only a few years of experience to reach the level of an experienced designer. "At present, the design of chip architecture depends on the experience of architects, if the experience accumulated by architects can be integrated into EDA tools with the help of AI technology, the threshold of chip design can be greatly reduced, and the efficiency can be greatly improved." Wang Bingda pointed out.
If it is AI that runs through the entire chip design process, developers need to have an understanding of how THE AI works. Wang Bingda explained, "Optimizing the process of chip design with AI technology requires customers to continuously adjust according to the actual progress. For example, in the traditional process, the time and order of each step are fixed, and the previous steps will be completed before the subsequent steps are entered. After adding AI, the time of step 1 may only need half of the original, and the time of step 2 only needs one-tenth of the original, at which time the user needs to adjust accordingly. ”
Of course, the integration of AI with EDA tools can not only significantly save R&D time, but also bring about improvements in chip performance and design costs.
Taking the DSO.ai of Synopsys Technology as an example, the leading IDM manufacturer in the United States has achieved remarkable results after adopting DSO.ai, the time of chip design has increased by 2-5 times, and the overall energy consumption of SoC chips has increased by 9%. Applying DSO.ai to the design process of different types of chips can result in significant time savings and performance gains with only one engineer.
"Different types and scenarios of chips, AI can bring about different improvements. This is because the entire design process of the chip needs to go through millions or tens of millions of steps, and the degree of improvement brought by AI in different processes is not consistent, and at the same time, the results of the previous step of optimization affect the effect of the next STEP of AI improvement. Wang Bingda pointed out, "After EDA joins AI, while saving chip design time, designers can focus on optimizing performance and innovation of core functions in the same time, which is naturally easier to design chips with better performance, and the overall cost can be reduced accordingly." ”
In the future, AI technology will be integrated into the whole process from the architecture design, manufacturing and packaging of chips. As for whether the cycle of chip design can change from year to month, Wang Bingda believes that it is clear to significantly shorten the cycle of chip design through AI + EDA, but shortening the entire life cycle of chips from design to manufacturing also requires the joint efforts of the entire industry chain.
3
The era of chip differentiation competition
Before further exploring the changes that AI will bring to the chip industry, it is necessary to answer a question. A key element of AI development is enough data, is there enough AI data to train EDA? Wang Bingda said: "EDA itself is a precision science, even before the arrival of AI, EDA has an accurate algorithm, the calculated data we call 'gold data'." The emergence of AI allows us to make better use of golden data training and make EDA tools more intelligent. ”
"EDA's AI doesn't rely as much on data as many industries, but it also needs user feedback to help us continue to improve the intelligence of EDA tools." Synopsys' unique strength is that we have the tools for the entire process of chip design, which allows us to use AI throughout the process, resulting in a more significant overall improvement. Wang Bingda further said.
When the user's design coincides with the trained tool, most of the design can be completed quickly, saving a lot of time, and the rest of the work is some optimization work.
"Users can also use the data they have to retrain EDA tools, so that customers can have more personalized and customized tools to design more distinctive products." Wang Bingda said, "Most of our products will open this interface. ”
But to better play the role of AI in chip design, how to find the integration point has become a challenge. "To give full play to the maximum benefits of AI in chip design, the difficulty lies in finding the most ingenious combination of AI and specific fields, which depends on the designer's cognition of the special field." Wang Bingda thinks.
In such competition, the advantages of system companies are more obvious. They know more about their own business, have a deeper understanding of algorithms, and have a lot of data, but they lack experience in chip design. But the EDA tool that integrates AI can just lower the threshold for system companies to design chips, and can also help them design chips faster and better.
"I believe that AI +EDA tools will soon be applied from digital design to almost all fields, and within a few years, there will be AI in all chip design processes." Wang Bingda said.
At that time, the competition in the chip industry may evolve into a competition between specialized chips in the field of system companies. How will General Chip face such competition?
Wang Bingda believes that the advantage of general chip companies lies in the understanding of chip architecture, and can make chips in the time window with the right process and the best cost, but what is lacking is an in-depth understanding of the system and terminal applications. Chip design companies need to find good system companies to work together to dig deeper into the needs to provide flexible, versatile chips that can accommodate multiple end applications.
In the face of these two types of customers, Synopsys provides completely different services. For system companies, the goal is to help them solve the choice of chip architecture and process through various IP modules and design tools; for general chip companies, the goal is to simulate the actual performance, power consumption and other performance before the chip is produced through faster and easier tools such as simulation verification and rapid prototyping, saving costs and design cycles.
END