Views: 0 Author: Site Editor Publish Time: 2024-11-07 Origin: Site
TrendForce, a global research firm, held a seminar on October 16, 2024, titled "AI Era, Semiconductor Global Expansion -2025 Technology Industry Big Forecast". The speech topics of this big prediction seminar cover high-tech industries such as wafer foundry, HBM, NAND Flash, AI PMIC, AI servers, panel level packaging, liquid cooling and heat dissipation, AI PC, and robotics.The rapid development of the artificial intelligence (AI) industry will undoubtedly become a powerful force in reshaping the industrial landscape, with wide-ranging and far-reaching impacts on the development dynamics of many other fields. TrendForce Consulting's Big Forecast delves into how AI can specifically impact the semiconductor industry and how it will lead new trends in the future development of the semiconductor industry.
Wafer foundry
Predicting the Development of AI Industry in 2025 from the Dynamics of Wafer Manufacturing
The demand for high-performance computing chips driven by AI applications has been hot for nearly two years, and high computing power applications have become the biggest driving force for advanced processes and the overall wafer foundry industry. Starting from 2025, in addition to AI chip suppliers and CSPs self-developed chips, memory suppliers are seeking to collaborate with advanced process wafer foundry partners to better adapt HBM and logic chips in order to meet high computing power demands. The upstream, midstream, and downstream supporting IP, design services, and back-end packaging and testing ecosystem of wafer fabs are all necessary resources for the AI arms race, no longer just focusing on advanced technologies in the front-end process.
From the overall analysis of the wafer foundry industry, in addition to the business opportunities of advanced processes, whether AI's demand for power management can bring new vitality to mature processes with long stable demand, and how the wafer foundry industry will transform under the development of Cloud AI and Edge AI in 2025, have become the focus of attention. Each application will gradually end its two-year inventory correction cycle in 2024. TrendForce predicts that the global wafer foundry industry will experience a 20% growth in output value by 2025, with TSMC still leading the way. Other wafer foundries are also expected to have an annual growth rate of nearly 12%.
Mature process
High computing power drives demand, AI PMIC injects new momentum into mature process development
Due to the ongoing uncertainty in the global economic outlook and the underperformance of China's domestic demand market, the main growth driver for the wafer foundry industry in 2024 will come from AI server related chips. As a result, the high capacity utilization rate of advanced processes is expected to continue until 2025. However, the recovery pace of mature processes above 28nm (inclusive) is relatively slow, and it is estimated that the average capacity utilization rate will only increase slightly by 5% to 10% by 2025, reaching a water level of around 80%; The average utilization rate of eight inch production capacity is around 75%, and there is an urgent need to seek new growth drivers to fill the capacity gap.
It is worth noting that as AI chip iterations continue to increase computing power, the corresponding thermal design power consumption (TDP) continues to increase. For example, the highest TDP of the Nvidia A100 is about 400W, which increases to 700W when entering the H100, and the next generation Blackwell series will break through the 1000W mark. The increasing TDP requires more Power ICs to assist in managing power transmission, reducing energy losses during conversion, and ultimately improving overall efficiency.TrendForce Consulting estimates that the number of Smart Power Stage (SPS) required for AI GPUs will rapidly increase with product iteration updates, driving related demand to grow 2-3 times between 2023 and 2025, becoming a new driving force to support mature process capacity.
Liquid cooling heat dissipation technology
Driven by AI, liquid cooling technology is heading towards a new blue ocean
In recent years, large American cloud providers such as Google, AWS, and Microsoft have actively built new data centers globally and accelerated the deployment of AI servers. Due to the upgrade of chip computing power, the thermal design power consumption (TDP) will significantly increase. For example, Nvidia's newly launched GB200 NVL72 cabinet solution will have a TDP of up to about 140kW, and a liquid cooling solution is required to effectively solve the heat dissipation problem. It is expected that the water to air (L2A) method will be the mainstream in the initial stage. It is estimated that by 2025, with the official mass shipment of GB200 cabinet solutions, the overall liquid cooling heat dissipation penetration rate of AI chips will increase from 11% in 2024 to 24% in 2025.
In addition, regarding the self-developed advanced AI ASICs by cloud industry players, Google is the most active American company in adopting liquid cooling solutions, while other cloud industry players still use air cooling as their main cooling solution. In addition, with the increasing awareness of ESG among global governments and regulatory agencies, it will accelerate the development of cooling solutions from gas cooling to liquid cooling. It is expected that the penetration rate of liquid cooling solutions will increase year by year, prompting power supply companies, cooling industry players, and system integration factories to compete in the AI liquid cooling market, forming a new industry competition situation.
AI server
Prediction of Key Development Trends in the Global AI Server Market and Industry Chain in 2025
Observing the development trends of the global AI market, in 2024, benefiting from the strong demand for building AI infrastructure from CSPs and brand customers, the annual growth rate of global AI server shipments (including GPU, FPGA, ASIC, etc.) is expected to reach 42%; By 2025, driven by demand from cloud providers and sovereign clouds, the shipment volume of AI servers is expected to grow by about 28%, driving the proportion of AI servers in the overall server market to nearly 15%. In the medium to long term, with the advancement of various cloud based AI training and inference application services, it is expected that by 2027, AI servers will account for nearly 19% of the overall server market.
Observing major AI chip suppliers, it is estimated that Nvidia's market share in the GPU market will approach 90% by 2024. It is estimated that the Blackwell new platform will gradually increase in volume and become the mainstream for NVIDIA's high-end GPU shipments in the first half of 2025. By adopting various solutions such as using pure GPUs (such as B200) or integrating our own Grace CPU GB200, we can meet the needs of different customer groups. Other industry players such as AMD, Intel, and CSPs are actively developing new generation AI chips, which will drive the growth momentum of shipments in 2025 and further double the growth of CoWoS and HBM shipments.
HBM
The demand challenges and future prospects faced by the HBM market
The HBM market is still in a high growth stage, and with the continuous deployment of AI servers, GPU computing power and memory capacity will be upgraded. HBM has become an indispensable key component, driving the improvement of HBM specifications and capacity. The Nvidia Blackwell platform will use 192GB HBM3e memory; AMD's MI325 has been upgraded to over 288GB. Due to the high production difficulty of HBM and significant room for improvement in yield, the overall production cost has been pushed up, with an average selling price of about three to five times that of DRAM products. With the mass production of HBM3e and the gradual expansion of production capacity, it is expected that its revenue contribution will increase quarterly.
NAND Flash
The NAND Flash Market Driven by AI: A Deep Analysis of Opportunities, Interweaving Challenges .
After experiencing huge losses in 2023, NAND Flash suppliers have become more conservative in their capital expenditures. At the same time, the demand for memory products such as DRAM and HBM will benefit from the AI wave, which will squeeze out equipment investment for NAND Flash in 2025, easing the severe oversupply situation in the past.With the rapid development of AI technology, the NAND Flash market is undergoing unprecedented changes. The increasing demand for high-speed and high-capacity storage in AI applications is driving the booming development of the Enterprise SSD (eSSD) market.
