Zhitong Financial APP learned that CSC released a research report that AIGC triggered a revolution in the content generation paradigm, ChatGPT led artificial intelligence applications into reality, GPT architecture was rapidly iterated, and cloud large model multi-modal became a development trend, bringing about a rapid increase in computing power resource consumption. Hardware infrastructure has become the cornerstone of development, requiring supporting upgrades in computing power, transportation capacity, storage capacity, heat dissipation and other fields, and the core benefits of computing power chips and other links.
FPGA, one of the computing power chip series: new fields open up new space, product power improvement accelerates localization
FPGA chip programmability provides greater flexibility, and technology iteration brings capacity and performance improvements. The functions of the FPGA chip are not fixed after the manufacturing is completed, and the user can configure the designed circuit through the special EDA software to configure the FPGA chip according to his needs, and convert it into a chip with specific functions, and it can be configured multiple times. FPGAs are more flexible than other chips, excel at massively parallel computing, and consume less power. The parallel computing performance of FPGA chips is provided by their capacity, and high-capacity FPGAs allow more processing circuitry to be deployed, resulting in higher processing performance. Technical iterations such as advanced processes, advanced packaging, and circuit optimization of basic component LUTs bring capacity and performance improvements. FPGA chips gain market share in various industries with their own flexibility, which puts forward high requirements for product definition and product line richness.
FPGA chips are widely used in downstream applications, and the industry has maintained rapid growth. With the acceleration of emerging markets such as data center construction, artificial intelligence and autonomous driving, FPGA demand will continue to grow. According to Gartner's forecast, the global FPGA market size will increase from $5.585 billion to $9.69 billion from 2020 to 2026, with a CAGR of 9.6%. According to Frost & Sullivan's forecast, the size of China's FPGA market will increase from 17.68 billion yuan to 33.22 billion yuan from 2021 to 2025, with a CAGR of 17.1%. From the perspective of downstream applications, communication and industry are the top two markets for FPGA chips, and the proportion of the communication market is expected to continue to increase driven by new technologies, and automobiles are the fastest growing downstream markets.
Industry concentration is high, medium and low capacity and mature process FPGAs meet the current market mainstream demand. The FPGA chip industry is highly concentrated, and the head companies occupy the vast majority of the market share. According to Gartner data, AMD, Intel, Lattice and Microchip accounted for 51%, 29%, 7% and 6% of the market in 2021, respectively, and the top four US companies accounted for 93% of the global FPGA market. At present, the main domestic FPGA manufacturers are Unigroup Guowei (including Unigroup Tongchuang), Fudan Microelectric, Anlu Technology, etc., occupying 16% of the domestic market in 2021.
Hashrate chip series: Chatgpt brings the outlook of computing power chip investment opportunities
AIGC has triggered a revolution in the content generation paradigm, ChatGPT has led artificial intelligence applications into reality, GPT architecture has been rapidly iterated, and cloud large model multi-modal has become a development trend, bringing about a rapid increase in computing power resource consumption. Hardware infrastructure has become the cornerstone of development, requiring supporting upgrades in computing power, transportation capacity, storage capacity, heat dissipation and other fields, and the core benefits of computing power chips and other links. Coupled with the restrictions on the supply of high-end GPU chips to China overseas, domestic manufacturers have ushered in a replacement window period, and the leaders in each link are expected to enter a stage of rapid development.
AIGC has triggered a paradigm revolution in content generation, and cloud algorithms have evolved to large model multimodalities. The proportion of inference in the computing power deployed in the cloud has gradually increased, indicating that the number of AI landing applications is increasing, and the release of ChatGPT has triggered a number of technology giants to carry out an AI arms race, or become a watershed in the maturity of artificial intelligence. The rapid iteration of GPT architecture and the increasing number of parameters drive higher and higher training accuracy, and the multi-modal development trend of large models in the cloud has become a development trend, bringing about a rapid increase in computing power resource consumption.
Hardware infrastructure has become the cornerstone of development, and computing power chips and other links have benefited from the core.
Computing power requirements, model training requires large-scale computing power chips to be deployed in intelligent servers, CPU is indispensable, but performance improvement has encountered bottlenecks, and the CPU+xPU heterogeneous solution has become the standard for large computing power scenarios. Among them, GPU parallel computing has obvious advantages, CPU+GPU has become the most popular heterogeneous computing system, while NPU has obvious performance and efficiency advantages in specific scenarios, and the application potential of the inference end is huge. In addition, in the post-Moore era, chiplet packaging is a cost-effective alternative to advanced processes and has become a trend in the semiconductor industry.
Storage requirements, huge training general data sets require corresponding storage hardware facilities, such as warm and cold storage, data access acceleration, data lakes and large-capacity storage, as well as storage protocols and access protocols specifically tailored for AI, and NVMe-oF protocols that support server-SSD communication are also expected to benefit from use.
Transportation capacity requirements, external access, internal data translation require high-speed network connection lines or switch systems, driving the upgrade of optical communication technology.
Heat dissipation requirements, AI server power consumption is relatively higher, the current mainstream heat dissipation solutions are evolving towards the chip level, chip liquid cooling market development potential is huge.
Domestic manufacturers ushered in the development window period, it is recommended to pay attention to the leading manufacturers in each link. NVIDIA and AMD are limited in supplying high-end GPU chips to China. Domestic computing power chips ushered in the window period of domestic substitution; Breaking through 4800TOPS and reducing deployment costs are the main points of domestic computing power chips. At present, a large number of domestic computing power chip manufacturers have emerged. The layout of the Chiplet industry chain is timely, focusing on the leading manufacturers in each link. High-speed interfaces such as PCIe and CXL indirectly benefit power chip manufacturers.
Attached: 1. Notable AIGC landing application: ChatGPT can replace the existing voice interaction model of smart speakers; 2. Other AI development directions: Nvidia DLSS, TinyML, in-memory computing, and solution integration based on advanced packaging.
Computing power chip series (3): advanced packaging investment opportunities in the era of large computing power
1. Application: Large computing power applications such as high-performance servers (HPC) and autonomous driving (ADAS) replace mobile phones/PCs as the driving force of a new round of semiconductor cycles, and the iteration of high-end packaging processes in the post-Moore's Law era has become a new development trend. In terms of TSMC's downstream applications, HPC's revenue growth rate has continued to lead since surpassing mobile phones in Q3 2020, and the corresponding revenue proportion surpassed mobile phones for the first time in Q1 2022 to become TSMC's downstream largest application, compared with only 16% of packaging and testing manufacturers in high-value computing electronics. CSC believes that with the increase in demand for large computing power, the replacement of advanced processes by advanced packaging has become the best solution to reduce the cost of unit computing power, thereby increasing the value of computing electronics in packaging and testing manufacturers.
2. Process: The 2.5D/3D packaging form represented by Chiplet has become the standard for large chips, and high-end packaging technologies such as TSV/RDL/Fan-out have increased the proportion of value in the packaging link. The growth of semiconductor value downstream from mobile phone / PC to high computing power HPC and ADAS transfer, packaging process began to Chiplet represented by 2.5D / 3D packaging transfer, from the packaging process flow, wafer foundry based on the advantages of manufacturing links extended to TSV process, packaging and testing plants are more involved in RDL and Fan-out and other packaging processes, with the expansion of the overall packaging and testing market of high computing power chips, packaging and testing manufacturers gradually expand the layout of 2.5D and 3D packaging and testing.
3. Market: TSMC, the global wafer foundry leader, has built a global benchmark for 2.5D/3D advanced packaging processes, and the growth of the packaging market in the next few years will mainly benefit from the expansion of advanced packaging. At present, the largest scale of advanced packaging revenue is wafer foundry leader TSMC, which is expected to contribute $5.3 billion in advanced packaging in 2022, and global packaging and testing leaders ASE and Amcos have launched 3D packaging and testing process platforms to actively seize the share of advanced packaging. It is expected that the size of the advanced packaging market will increase to $65.1 billion in 2027, with a CAGR of 9.6% from 2021 to 2027, and advanced packaging will become a new growth momentum for packaging manufacturers in the era of large computing power.
In-depth research on the GPU industry: The wave of AI large models is rising, and GPU chips are standing at the forefront again
The GPU has two core functions: graphics rendering and parallel computing. GPUs have a large number of arithmetic units, suitable for computationally intensive and easily parallelized programs, and are generally responsible for graphics rendering and parallel computing as coprocessors. GPU microarchitecture is composed of stream processor, texture mapping unit, rasterization processing unit, ray tracing core, tensor core, cache and other components, the design of microarchitecture plays a vital role in improving GPU performance, and is also the most critical technical barrier in the GPU research and development process. In terms of parallel computing, the birth of CUDA (Unified Computing Device Architecture) greatly reduces the programming difficulty of GPGPU parallel computing, realizes the generalization of GPUs, and "personal computers" become "supercomputers" that can realize parallel computing.
The global GPU market has maintained good growth, and AI servers have become the core support for market growth. The global market size of GPUs is expected to be $59.5 billion in 2023, and the industry will maintain rapid growth, with a CAGR of 32.9%. The market players of GPU can be divided into personal computer GPUs, server GPUs, autonomous driving GPUs, and mobile GPUs. Over the past few quarters, the PC GPU market has suffered a huge hit, with shipments falling significantly. There are three core reasons:
First, the personal computer market is in a downward cycle;
Second, the ebb of virtual currency mining has caused a huge impact on independent GPU shipments;
Third, downstream board manufacturers start the inventory reduction cycle. Recently, various adverse factors are gradually dissolving, and the personal computer GPU market has ushered in the dawn. Server GPUs are mainly used for AI and high-performance computing, and the rapid development of the artificial intelligence industry has brought strong demand for AI computing power, and AI servers have become the core support for the growth of the GPU market. The natural language big model represented by ChatGPT shows a high degree of intelligence, and the generative AI capability continues to break through to the stage of auxiliary productivity, and the demand for AI model computing power has reached a new level, bringing a significant pulling effect to the server GPU market. Autonomous driving GPUs have significant technical advantages in high-level autonomous driving, and with the gradual increase in the penetration rate of high-level autonomous driving, the autonomous driving GPU market has also entered a stage of rapid growth.
With the core technology advantages of its data center GPUs, NVIDIA has become a global leader in artificial intelligence chips. Nvidia used to focus on GPU chip design, and has transformed into a computing platform company and a leader in artificial intelligence chips. Its main businesses include gaming & entertainment, data center, professional visualization, and automotive business. In the past two quarters, NVIDIA's gaming & entertainment revenue has declined sharply along with weak overall market demand for PC GPUs.
As the impact of virtual currency mining on GPU discrete graphics cards gradually declines, the company's 22Q4 game & entertainment business has increased again month-on-month, and CSC believes that the company's personal computer GPU business is gradually returning to a normal growth stage. In 2022, NVIDIA's data center business revenue surpassed the game & entertainment business to become the largest source of revenue, the company's GPGPU has core technology advantages, dominates the AI chip market, and its data center business will contribute long-term power to the company's high-quality growth. In terms of autonomous driving business, NVIDIA provides full-stack autonomous driving solutions, at the hardware level, its Orin and Thor autonomous driving chips provide large computing power, while DLA modules and PVA modules realize AI algorithm acceleration; Providing a complete developer kit at the software level, the platform-based advantages of its autonomous driving business ensure NVIDIA's leading position in high-level autonomous driving.
AMD, as the world's leading chip designer, competes with NVIDIA in the GPU market. AMD's data center business and embedded business show good growth, and the company offers both personal computer GPUs and data center GPUs. The company's integrated GPUs are mainly used in desktop and notebook APU products, which have more cost-effective advantages than discrete GPUs. The Radeon series of discrete GPUs is built on the RDNA 3 architecture and uses a 5nm process and chiplet design to achieve an overall improvement in performance. AMD introduced the Radeon Instinct GPU-accelerated chip for data centers, and the Instinct series is based on the CDNA architecture. The latest CDNA 2 architecture has significantly improved computing power and interconnection capabilities, and the performance indicators of the computing chip MI250X using the CDNA 2 architecture are the same as NVIDIA's advanced computing chip. AMD ROCm benchmarks NVIDIA CUDA, and its computing ecosystem is also constantly enriched.
Major players on mobile include Qualcomm, ARM, and Imagination. Mobile GPUs are limited in terms of energy consumption and volume during the design process, and they all appear on the mobile side in the form of integrated SOC chips. Qualcomm maintains a leading position in the flagship Android smartphone SoC market, ARM is a leading GPU IP company, Imagination's PowerVR architecture has been widely recognized by the market in the field of mobile chips, and then successively proposed an upgraded version of PowerVR's IMG series architecture.
The domestic GPU market is vast, and a number of excellent GPU designers and manufacturers have emerged. According to Verified Market Research, the Chinese mainland GPU market size was $4.739 billion in 2020, and it is expected that the Chinese GPU market will reach $11.1 billion in 2023. With the recent macroeconomic recovery and the increase in AI computing power of domestic Internet companies, the rising demand for PCs and servers is expected to bring an overall pulling effect to the domestic GPU market. A number of excellent GPU designers and manufacturers have emerged in China, such as Haiguang Information and Jingjiawei. The product performance of Haiguang Information DCU has reached the international mainstream high-end processor level of the same type, and is in a leading position in China, and the Haiguang Information DCU coprocessor is fully compatible with ROCm GPU computing ecology. Jingjia Micro GPU has a long history of research and development, deep technology accumulation, superior GPU performance, and the overall chip business shows a good growth momentum.
Dynamic tracking of the global power chip leader (1): The silicon carbide business has entered an explosive period
Look at the industry development trend from the overseas power chip leader in 2022 mid-year report
1. Overseas power leading manufacturers have revised their 2022 revenue guidelines, Q3 outlook for an average growth of 33% From the performance of Infineon, ON Semiconductor and STMicroelectronics, Q2 revenue increased by 33%, 25% and 28% year-on-year, respectively, of which Infineon raised its 2022 revenue target from 13.5 billion euros to 14 billion euros, and ON Semiconductor's Q2 quarterly revenue exceeded the upper limit of the Q1 guidance range. ST raised its full-year revenue guidance to $15.9-$16.2 billion from $14.8-15.3 billion. Looking forward to the third quarter, the prosperity of the automotive semiconductor field continued, with month-on-month growth of 8%, 2% and 11%, respectively, and the year-on-year growth rate averaged 33%.
2. The penetration rate of electric vehicles continues to increase, leading the growth of automotive semiconductor demand, and the progress of silicon carbide MOSFETs has accelerated into an explosive growth period divided according to downstream application fields, with Infineon, ON Semiconductor and STMicroelectronics accounting for 47%, 38% and 38% of automotive semiconductor revenue, respectively, and the year-on-year growth of 41%, 41% and 35% in Q2 quarters, far exceeding the company's overall revenue growth. In addition, the three leading manufacturers continue to grow orders in hand, taking Infineon as an example, 50% of the new orders of 5 billion euros come from automotive customers. In the automotive semiconductor field, silicon carbide chips are more prosperous, and ON Semiconductor raised the company's full-year silicon carbide revenue target for 2022 to three times that of 2021, while guiding the company's silicon carbide revenue to exceed the $1 billion mark in 2023, reflecting that the industry is in a period of explosive growth.
3, overseas power IDM manufacturers actively layout 12-inch production line, optimistic about the cost advantage brought by the production line upgrade In addition to the demand side of the automotive and industrial demand continues to grow, the supply side is also undergoing obvious changes, relative to the logic chip foundry feedback began to enter the destocking stage, power IDM manufacturers in the second half of this year annual capacity supply is still relatively tight, another very important trend on the supply side is the power IDM company production line from 8 inches to 12 inches upgraded.
Analog chips have long slopes and thick snow, and local manufacturers have accelerated their growth
1. The space of analog chips is vast, and the demand for communications, industrial control and automotive drives the growth of the industry.
According to IC Insights, analog IC market sales in 2021 were $74.1 billion, up 30% year-on-year, and analog chip sales are expected to grow 12% to $83.2 billion in 2022. From the perspective of market space, general-purpose chips account for about 40% of the analog chip market, and special-purpose chips account for about 60%. Communications, industrial control and automotive are the main downstream demand sources for analog ICs, which are expected to account for 81.7% in 2022. Communication (including mobile phones) is still the largest market for analog ICs, and the increase mainly comes from 5G infrastructure and terminals, and satellite demand is gradually emerging. Vehicle electrification and intelligence drive demand growth, and the power domain brings more incremental markets. The demand for analog ICs in the industrial market is fragmented, and applications such as industrial control, robotics, and power are driving growth. In consumer appliances, the domestic production rate of home appliance chips has great room for improvement, and the future market space for wearable devices is broad.
2. The localization rate of analog chips is low, and the low industry concentration brings opportunities to domestic manufacturers. According to IC Insights, the Chinese market will be $29.4 billion in 2021, and the compound growth rate of China's analog IC market is expected to be 7.85% from 2021 to 2026. In 2021 Chinese mainland total revenue of Top10 analog IC manufacturers will be about 2.286 billion US dollars, accounting for 7.78% of China's analog IC market share, and the localization rate is still at a low level. Due to the diversity of analog IC product types, the market share of industry leader TI is 19%, which is much lower than the market share of digital chip leading manufacturers, giving domestic manufacturers greater market opportunities. Domestic substitution is timely, mergers and acquisitions, integration and transformation of IDM are the general trend, local manufacturers focus on subdivision to form advantages, and are expected to grow into platform-level manufacturers.
3. The industry has shifted from supply-driven to demand-driven, and the localization process continues to advance. 2021 is a year of severe global semiconductor component shortages, and production capacity determines performance. Due to the widespread use of mature processes for analog ICs, the industry expansion progress is relatively slow, with the gradual release of TI's new production capacity in the second half of 2022 and early 2023, the industry's production capacity will be structurally alleviated, and future demand will determine growth. In the medium and long term, the analog IC industry will maintain a rapid growth rate, and the localization rate will maintain a rapid improvement momentum, and the head concentration effect will be more obvious.
Memory chips: long-term high-growth track, local manufacturers are expected to rise
1. The largest segment of semiconductors, 5G, AI, and intelligent vehicles drive growth. Memory chips are an important branch of semiconductors, accounting for 30% of its output value, and the global market size has maintained an upward trend in fluctuations, increasing from US$54.6 billion in 2005 to US$122.9 billion in 2020, with a compound growth rate of 5.6% in the past 15 years. Looking back at history, the memory chip market has seen many rounds of new terminal or application-driven growth cycles, such as the penetration of PCs in the 90s, the penetration of functional machines and the launch of iPods in 2000, the penetration of smart phones and the outbreak of cloud computing in the 2010s, and the future demand for memory chips will enter the next round of growth cycle driven by 5G, AI and smart cars. The market size is expected to grow by 22% year-on-year in 2021 and exceed $200 billion in 2023.
2. Short-term strong cycle, long-term high growth. Driven by downstream demand such as servers, mobile phones, and PCs, the scale of the memory chip market has expanded rapidly. CSC judged that in the next few years, servers, mobile phones, and PCs will still be the core applications of memory chips, of which server demand growth is obvious and will become the main driving force, while the demand growth of mobile phones and PCs mainly comes from stock upgrades. In addition, benefiting from automotive intelligence, automotive storage demand will continue to increase, and the long-term market is expected to be comparable to PC. From the structural point of view, DRAM and NAND are the core categories of memory chips, and their total share in 2020 will reach 96%, of which DRAM technology iteration is mainly process evolution, and NAND technology iteration is mainly 3D stacking. The memory chip market pattern tends to be oligopoly, so the mismatch between the supply of large manufacturers and downstream demand has led to the storage cycle, short-term dimension, memory chip price cyclical fluctuations, long-term dimension, memory chip unit capacity price into a downward trend.
3. China's memory chip self-sufficiency is insufficient, and local manufacturers are expected to rise. At present, the total market share of the top three manufacturers in the DRAM market is greater than 95%, the pattern is highly monopolized by Samsung Electronics, Micron, SK Hynix, the Matthew effect is significant, and NAND is relatively scattered, CSC judges that its pattern will evolve to DRAM in the future. The demand for domestic memory chips is huge, the market size is more than 43 billion US dollars, more than 1/3 of the world, and the growth rate is higher than the global average, but the self-sufficiency rate is less than 5%, and the huge domestic demand, emerging applications and policies promote the rapid development of domestic memory chips. From the downstream point of view, key areas such as government, telecommunications, and finance will become the backbone of the localization of memory chips, with potential demand exceeding 10 billion US dollars.
4. Investment advice: storage is a long-term high-growth track, the domestic demand is huge, the localization demand is urgent, excellent manufacturers have emerged in the bulk market and niche market, and CSC is optimistic about the rise of local manufacturers for a long time.