TrendForce forecasts that NVIDIA will focus on promoting its Blackwell B300 series AI chips in 2025, with increased adoption of CoWoS-L packaging and liquid cooling solutions. The B300 series will replace Blackwell Ultra, with shipments expected to commence between Q2 and Q3 2025, primarily targeting performance-focused CSP customers and driving the development of HBM3e 12hi memory. Furthermore, NVIDIA will continue to promote its NVL72 Rack solution, further accelerating the adoption of liquid cooling in AI servers. Improvements in CoWoS-L production capacity and yield rates will also contribute to its wider adoption in AI chips.
TrendForce indicates that while Samsung, SK Hynix, and Micron have all sampled HBM3e 12hi products, challenges remain in yield ramp-up and qualification, leaving uncertainty over the risk of HBM oversupply in 2025. Despite aggressive TSV capacity expansion by DRAM suppliers, the mass production learning curve for HBM3e 12hi is long, and AI chips demand higher stability, creating uncertainty in actual capacity release. TrendForce maintains its forecast of HBM accounting for 10% of total DRAM bit output in 2025 and contributing over 30% of DRAM market value. HBM3e market share will be a key differentiator in DRAM suppliers' profitability.
TrendForce indicates that DDR4 and LPDDR4X prices are facing downward pressure due to sluggish consumer demand and CXMT's capacity expansion. Demand for DDR4 in PC, server, and consumer electronics markets is declining, while the smartphone market is experiencing an oversupply of LPDDR4X, mainly driven by CXMT's expansion. Samsung plans to cut LPDDR4X production and reduce prices to address the competition. The DRAM market is expected to face dual pressures of shrinking demand and oversupply in Q4, potentially leading to continued price declines.
TrendForce forecasts that the AI server liquid cooling market will experience rapid growth, driven by the launch of NVIDIA's Blackwell platform and ASIC upgrades, with penetration rate projected to reach 20-30% by 2025. Blackwell's high performance and power consumption will accelerate the adoption of liquid cooling solutions, prompting ODMs, power supply, and cooling solution providers to actively expand into the liquid cooling market. Additionally, the rising global awareness of ESG and CSPs' investments in AI servers will further promote liquid cooling technology. Due to trade conflicts, cooling suppliers are accelerating expansion in Southeast Asia to seize growth opportunities in the AI accelerator and liquid cooling markets.
(a) incentives for chip makers to adopt FOPLP solution: AI GPU makers pursuing chip packaging size expansion could adopt chip-last solution, while power IC and RF IC makers pursuing packaging cost reduction could adopt chip-first solution. The segmented market could accommodate different types of suppliers.
(b) development of FOPLP solution by foundry: expanding chip packaging size of 2.5D packaging for AI GPU from wafer level to panel level.
(c) development of FOPLP solution by OSATs: migrating from power IC and RF IC to CPU and GPU, enabling multi-chip packaging.
(d) development of FOPLP solution by panel makers: entry into chip packaging business, supported by licencers.
This report explores the impact of US semiconductor export controls on China’s semiconductor industry and analyzes why Chinese manufacturers have more room to maneuver in the mature process segment. This report also highlights the potential for electronic design automation (EDA) software to expand import substitution and focuses on key suppliers in this area.
1. An Overview of the Semiconductor Market
2. Demand & Capacity Utilization
3. Supply & Geopolitical Impact
4. The Semiconductor Market in the Age of AI
5. TRI’s View
1. New Era for Japanese Semiconductor Industry: Innovations and Development Trends under Governmental Interventions
2. TRI's View
As indicated by the survey of global market intelligence firm TrendForce, chip suppliers such as AMD have been actively involved in negotiations with TSMC and OSAT throughout 2Q24 on chip packaging using the FOPLP technology, with major partnership models and items shown below:
Model 1: OSATs transitioning from traditional packaging method to FOPLP when packaging consumer ICs: AMD is currently negotiating with PTI and ASE on PC CPUs, whereas Qualcomm is talking it over with ASE on PMIC. Prior to this wave of demand, MediaTek had been worked with PTI on PMIC, while Qualcomm was already a partner of ASE pertaining to PMIC.
Model 2: foundries and OSATs transitioning from 2.5D wafer level to panel level when packaging AI GPUs: AMD and NVIDIA are currently negotiating with TSMC and SPIL on packaging AI GPUs by switching from the existing 2.5D wafer level to panel level, followed by an expansion in chip packaging sizes.
Model 3: panel providers packaging consumer ICs: NXP and STMicroelectronics are negotiating with Innolux on PMIC products.
Optical module is an indispensable component for achieving optical communication. However, corresponding issues are becoming increasingly apparent alongside the continuously climbing usage and growing transmission performance. This report, aside from probing into the development bottlenecks of existing optical module solutions, also focuses on the dynamics of three major solutions - LPO, OBO, and CPO - and their supply chains.