Artificial intelligence enters chip design and uses technology to feed back technology

arts/

At present, the chip has become the soul of the digital age, and it is also one of the three elements of the information industry. Chips are inseparable from washing machines, mobile phones, computers, etc., as small as daily life, or as large as various CNC machine tools in traditional industries and missiles, satellites, rockets, and warships in the defense industry. The chip market has also grown from $33 billion in 1987 to $433 billion in 2020.

However, at present, the latest trend of chip development seems to be difficult to follow Moore's Law: "The number of transistors that can be accommodated on an integrated circuit will double approximately every 18 months, and the performance will also double." With the increase of people's requirements for computing power, the development of chips has become more and more difficult, especially in chip design. In this context, artificial intelligence "feedback" chip design has become a new way out for chip design, which is also of great significance to the development of our country.

The importance of chip design

From the perspective of the core industry chain, although the chip industry is complicated and cumbersome, it cannot be separated from three steps: design, manufacturing, packaging and testing. Among them, the core of the core strength of the chip lies in the chip design.

You must know that for chips, the design and process are equally complicated. The birth of EDA technology in the 1980s-chip automation design greatly reduces the difficulty of chip design and VLSI. Engineers only need to use chip design for the function of the chip. The language description is input into the computer, and then the EDA tool software compiles the language into a logic circuit, and then debugs it. Just as editing documents requires Microsoft Office and photo editing requires Photoshop, chip developers use the EDA software platform for circuit design, performance analysis, and generation of chip circuit layouts.

Specifically, to design a chip, the developer must first clarify the requirements, determine the "specification" of the chip, define key information such as instruction set, function, input and output pins, performance and power consumption, and divide the circuit into multiple small modules. , clearly describing the requirements for each module. Then the "front-end" developers design the "circuit" according to the function of each module, use the computer language to build the model and verify that its function is correct. "Back-end" developers have to design a "layout" according to the circuit, and regularly reprint hundreds of millions of circuits onto a silicon wafer according to their connection relationships. At this point, the chip design is complete.

At the same time, many variables need to be considered, such as signal interference, heat distribution, etc., and the physical characteristics of the chip, such as magnetic field, signal interference, are very different under different processes. There is no mathematical formula that can be directly calculated, and there is no Applicable empirical data can be directly filled in, and we can only rely on EDA tools for step-by-step design, step-by-step simulation, and constant selection.

Of course, the verification after the design is not a simple thing. The goal of chip verification is to iteratively verify through inspection, simulation, prototype platform and other means before chip manufacturing, to discover system software and hardware functional errors in advance, optimize performance and power consumption, and make the design accurate and reliable, and meet the originally planned chip specifications. After each simulation, if the effect is not satisfactory, it must be redesigned, which is a great test for the wisdom, energy and patience of the chip developers.

If the architecture design and verification of the chip are still at the technical level, then the tape-out will directly enter the stage of "burning money" . Tape-out is trial production. After the design is completed, some small batches are produced by the chip foundry for testing. According to the market quotation, taking the Qualcomm Snapdragon 855, the processor with the smallest bare core area in the industry, as an example, the standard price for a 28-nanometer process is 499,072.5 euros, or nearly 4 million yuan.

The most important thing is that such a complex design cannot have any flaws, otherwise it cannot be repaired, and it must be done all over again . However, if it is redesigned and processed, it generally takes at least one year, and then tens of millions of dollars are invested, and sometimes hundreds of millions are required. That's why chips often take years to design.

However, if the chip design in the past still catches up with Moore's Law, then now, with the advancement of artificial intelligence technology, people's requirements for computing power are getting higher and higher, and the changes in this demand are recorded in weeks or days. In contrast, chip design takes significantly longer, which means that the design speed of new microprocessors can no longer meet the iterative development of algorithms, which creates a mismatch between supply and demand.

Moore's Law is failing, and how to shorten the chip design cycle has become an urgent problem for the semiconductor industry.

Artificial intelligence feeds back chip design

In fact, although chip design is a complex and lengthy process, in terms of decomposition, chip design is mainly composed of two main elements: layout and wiring. Wires are used to connect components virtually .

Taking chip layout as an example, the reason why chip layout is complex and time-consuming is because the process involves logic and memory modules, or the cluster setting needs to take into account power consumption, performance, area, etc., applications, IP and other components may restrict the chip. Design factors, engineers' familiarity with different tools and methodologies also vary.

For example, the same design goal can be achieved either with larger processors for higher performance, or with smaller, more specialized processing elements that are more tightly coupled to software. So even within the same domain and the same power design goals, there are many different ways to achieve the same goal. And the evaluation criteria for the pros and cons of the program also vary according to the specific needs of the field and the supplier.

In the past, when Moore's Law was in effect, the process only needed to be fine-tuned for the actual situation - as the chip iterated, the number of transistors had increased from thousands to billions, which made the design of the transistor arrangement on the chip. Heterogeneity is getting higher and higher. Instead of just thinking about how to arrange more transistors in the same space, chip designs now also need to take into account complex factors such as power density, thermal budget requirements, various types of mechanical and electrical stress, proximity effects, and the operating environment. This greatly increases the time-consuming design process, and also increases the design cost.

In addition, the optimization process has become more complex due to the ever-increasing demands on chip security. Depending on the importance of the device usage scenario, its security requirements vary. At the same time, the chip layout also needs to comply with the principles of wiring density and interconnection.

So, to unravel these factors, EDA vendors are turning to artificial intelligence and machine learning technologies for help. In fact, 90% of the wiring process in chip design has been automated, and only the last 10% of the work needs to be done manually, and the participation of artificial intelligence can further shorten the last 10% of the time. In essence, whether it is human intelligence or artificial intelligence, the purpose is to achieve chip optimization, but artificial intelligence is obviously more efficient in the process .

The "intelligence" of artificial intelligence comes from the optimal solution in different situations obtained by a large number of trials and policy adjustments in the data set. In the new scenarios encountered in actual production, AI matches these optimal solution strategies with the actual situation, so as to obtain the optimal answer relative to the actual scenario. In addition, artificial intelligence can also use reinforcement learning methods (RL) to guide training results.

Specifically, AI can model chip layout as a reinforcement learning problem, and the goal of reinforcement learning systems is to reduce power, improve performance, and reduce area. In April 2020, a team at Google published research on automatic layout using reinforcement learning (RL). When designing circuit layouts, the benefits of RL algorithms can be found in using fewer wires, more efficient use of space, or lower power consumption. In initial experiments, the algorithm was implemented in 24 hours compared to a solution that a human designer could find in 6 to 8 weeks, and the algorithm was successfully designed to reduce the overall wiring required for the chip, thereby increasing efficiency.

It's fair to say that the introduction of artificial intelligence into the chip design process to improve efficiency is now a general trend, at least for the major chip suppliers.

Chinese chips are catching up

In fact, today's semiconductor industry chain is actually a monopoly model dominated by the United States . The U.S. semiconductor industry accounts for almost half of the global market share, although in the 1980s the U.S. semiconductor industry suffered significant losses in global market share. In the early 1980s, U.S.-based producers accounted for more than 50 percent of global semiconductor sales. But due to intense competition from Japanese companies, the effects of illegal "dumping" and a severe industrial recession from 1985 to 1986, the U.S. semiconductor industry lost 19 global market shares and ceded global market share leadership to Japan.

But over the next 10 years, the U.S. semiconductor industry began to rebound, and by 1997 it had regained its leadership with more than 50% of the global market share, a position it holds today . American semiconductor companies maintained their competitive advantage in microprocessors and other leading equipment and continued to lead in other product areas. In addition, U.S. semiconductor companies maintain leadership in R&D, design and process technology.

Among them, the revenue gap of the top five chip design giants in the world has already appeared - the four chip giants of Qualcomm, Broadcom, MediaTek and Nvidia have entered the top five chip design companies for ten consecutive years , and AMD has ranked in the top five for seven times in ten years. Qualcomm, Broadcom, Nvidia and AMD are all American companies.

Among the five chip giants, Qualcomm is the leader in smartphone SoC and RF front-end, and also has a large number of communication patents; NVIDIA is the absolute leader in the global GPU market, with revenue of US$26.91 billion in fiscal year 2022, making it the world's second largest chip design company. ; Broadcom is a veteran semiconductor giant in the United States, with a high market share in various semiconductor products such as set-top box SoCs, wired network chips, radio frequency front-ends, Wi-Fi chips, and corresponding software services; MediaTek is Qualcomm's main competitor. Mobile phone SoC, TWS headset chip, IoT chip and other fields have layouts.

According to IC Insights ranking, in 2012, the top ten chip design companies in the world were Qualcomm, Broadcom, AMD, NVIDIA, MediaTek, Meiman Electronics, Isahua, Xilinx (acquired by AMD in 2022), Altera (acquired by AMD in 2015) Acquired by Intel in 2000) and Avago.

At that time, the revenue of fifth-place MediaTek and sixth-place Meiman Electronics were US$3.3 billion and US$3.1 billion respectively, a difference of only US$200 million. In 2021, the revenue gap between fifth place AMD and sixth place Novatek has widened to $11.6 billion , sixth place Novatek + seventh Meiman Electronics + eighth Realtek + ninth Xilin The sum of Si's annual revenue is only $16.5 billion, which is $100 million more than AMD.

In addition, in the context of the epidemic, in order to ensure their own technical strength and product competitiveness, these chip design giants are still developing new product lines: Qualcomm has begun to strengthen automotive and AR/VR businesses; NVIDIA has pushed GPU+CPU+DPU data center strategy , but also to acquire Arm; AMD acquires Xilinx to strengthen FPGA, etc.

In contrast, the overall volume of domestic chip design is still small, and the revenue gap between chip design companies and major global benchmarking companies is relatively large. Most companies are less than 5% of the revenue of benchmarking companies. The revenue of leading chip design companies in the field is basically at the level of tens of billions of dollars. Relevant companies mainly include Huawei's HiSilicon, UNISOC, Beijing Haowei, and ZTE Microelectronics.

After all, compared with the current advanced semiconductor industry in the United States, China is still catching up rapidly in the semiconductor industry. After all, when the United States establishes its global leadership in semiconductors from various perspectives such as legislation, industrial policy, direct intervention, and trade wars, China is deeply involved in the movement. The world semiconductor industry is changing with each passing day, and when China returns to the world again, it will be 20-30 years behind .

In front of this chasm that has grown over the years, even though China has spent tens of billions of dollars trying to stand out in the competition for semiconductors, faster computers and smartphones, and more sophisticated equipment, for now, China's The road to semiconductor catch-up is still a long way off. It is an indisputable fact that the localization rate of core chips is low.

Fortunately, although China lags behind the United States for many years in chip technology, in reality, relying on national strength and unified direction, it is still a huge advantage that China cannot match.

On the one hand, in the top-level design to support the development of the semiconductor industry, we must not only invest in capital and talents, but more importantly, we must reform the system and mechanism . Support high-end scientific and technological talents to carry out industry-university-research cooperation, or support high-end scientific research talents to bring technology to start a business, allowing them to conduct trial and error in a more market-oriented way. On the other hand, it is necessary to grasp the technical context. As mentioned earlier, there are many fields that have shown the advantages of artificial intelligence, and artificial intelligence will show this advantage in more fields in the future.

Standing in an era of technology, responding to technology with technology, and cultivating deep into the underlying technology, may be the road sign that should be erected on this road.