Workflow
HBM内存
icon
Search documents
存储芯片涨价潮下的国产替代机遇
Hu Xiu· 2025-09-30 06:04
根据最新市场动态,2025 年第三季度存储芯片价格呈现全面上涨态势,不同品类涨幅差异显著,部分产品价格已接近历史高位。 内存(DRAM)方面,DDR4内存受厂商减产和需求刚性影响,价格涨幅最为突出。32GB DDR4 内存价格从年初的不足 300 元涨至 500 元以上,半年内涨 幅超 66%。HBM内存作为 AI 芯片的核心组件,价格涨幅最为惊人,HBM2e 价格从 2024 年第三季度的 25 美元 / GB 涨至 2025 年第二季度的 45 美元 / GB, 半年涨幅达 80%。HBM3e 价格更是突破 100 美元 / GB,成为存储市场的 "价格标杆"。 存储卡(NAND)方面,消费级 SSD 硬盘价格在一个半月内上涨40%,例如 350 元的 1TB 硬盘涨至 550 元。企业级高容量 SSD硬盘(如 8TB 以上)因数据 中心需求激增,价格涨幅超过50%,例如采用 3D NAND 堆叠技术的产品需求旺盛,16TB SSD 价格从年初的 2500 元涨至 3800 元,涨幅 52%。 二、存储芯片国产替代的"双子星" 一、为什么近期存储芯片的涨幅这么大? 供需失衡是涨价的核心!最近1年,AI 服 ...
存储巨头,纷纷投靠台积电
半导体芯闻· 2025-09-24 10:47
如果您希望可以时常见面,欢迎标星收藏哦~ 来源:内容来自半导体芯闻综合。 美光科技,作为美国最大的电脑内存芯片制造商,因人工智能设备需求带动,给出了乐观的季度业 绩预期,其指出,下一代HBM内存将与台积电合作生产基础逻辑芯片。 公司周二发布声明称,本财年第一季度营收预计约为125亿美元,高于分析师平均预期的119亿美 元。扣除部分项目后,每股利润预计约为3.75美元,而市场此前预估为3.05美元。 这一前景证明了美光已成为AI投资的主要受益者。其高带宽存储器(HBM)是开发人工智能模型 所需芯片与系统的关键组成部分,这也使得该技术成为总部位于爱达荷州博伊西的美光极具盈利性 的产品。 值得关注的是,美光在财报会议上表示其HBM4 12-Hi DRAM 解决方案进展顺利,为了满足近期 不断增长的性能需求,该公司已生产并交付了迄今为止速度最快的 HBM4 解决方案的首批样品, 该解决方案提供超过 11 Gbps 的引脚速度和 2.8 TB/s 的带宽。他们表示,新的 HBM4 产品在性 能和效率方面应该超越所有竞争对手。 除了 HBM4,美光还谈到了下一代 HBM4E 内存。与完全基于自主研发的先进 CMOS 基础 ...
HBM,碰壁了
半导体行业观察· 2025-09-13 02:48
公众号记得加星标⭐️,第一时间看推送不会错过。 曾经炙手可热的HBM内存,似乎在一夜之间黯然失色。 最新发布的英伟达Rubin CPX GPU——一款专门针对预填充阶段优化的芯片,出人意料地选择了成 本更为亲民的GDDR7内存,而非业界习以为常的高端HBM方案。瞬间在行业内引发一场热议。 回顾过往数年,英伟达的AI芯片几乎无一例外地搭载最新一代HBM内存,以满足AI训练和推理对超 高内存带宽的苛刻要求。如今Rubin CPX GPU却反其道而行之,转向带宽相对较低的GDDR7,这不 禁让人产生疑问: 曾经的"内存之王"HBM,真的 会迎来新的威胁吗? 放弃HBM的AI芯片 至于落地时间表,英伟达计划在 2026 年随 Vera Rubin NVL144 CPX 机架一起推出 Rubin CPX GPU。该机架的配置堪称"巨兽":144 块 Rubin GPU + 144 块 Rubin CPX GPU,36 颗 Vera CPU,100 TB 高速内存,1.7 PB/s 内存带宽。 在这种组合下,整体性能将达到 8 ExaFLOPs NVFP4,是现有一代 GB300 NVL72 的 7.5 倍,也超 过未配 ...
【大涨解读】内存:海外大厂连续涨价,国产存储龙头或将跟进,英伟达最新GPU也转向GDDR内存
Xuan Gu Bao· 2025-09-12 02:46
Market Overview - On September 12, the storage sector experienced a significant opening surge, with companies like Deminor hitting the daily limit, and others such as Dongxin Co., Jiangbolong, and others seeing substantial increases [1] Company Insights - **Xingsen Technology (002436.SZ)**: A pioneer in the domestic IC packaging substrate industry, with applications in mobile memory modules [3] - **Deminor (001309.SZ)**: Transitioning from losses and planning to reduce holdings; has established a comprehensive storage product matrix including mobile storage and solid-state drives [3] - **Dongxin Co. (688110.SS)**: One of the few companies in mainland China providing a full range of storage chips including NAND Flash, NOR Flash, and DRAM [3] - **Jiangbolong (301308.SZ)**: A leading domestic storage module company, with memory products covering mainstream types like LDDR memory modules [3] - **Beiyi Innovation (603986.SS)**: A leader in the storage field, focusing on the development and sales of memory chips and microcontrollers, ranking first globally in non-foundry Flash supply [3] Industry Trends - Major US storage leaders like Micron and SanDisk saw collective gains, with SanDisk announcing a 10% price increase across all channels for consumer products, indicating a potential new pricing cycle due to changing supply-demand dynamics [5][6] - The global DRAM market size increased approximately 17% quarter-on-quarter to $30.9 billion in Q2 2025, driven by the rise of generative AI and increasing demand for DRAM contracts and HBM shipments [6] - Nvidia's launch of the RubinCPX GPU aims to enhance speed in handling massive data, utilizing GDDR7 memory, which could lead to increased demand for DRAM [5][6] - The establishment of Changchun's third phase company marks a significant step in China's semiconductor industry, potentially breaking the monopoly of Samsung and SK Hynix in the NAND market [6]
英伟达Rubin CPX 的产业链逻辑
傅里叶的猫· 2025-09-11 15:50
Core Viewpoint - The article discusses the significance of Nvidia's Rubin CPX, highlighting its tailored design for AI model inference, particularly addressing the inefficiencies in hardware utilization during the prefill and decode stages of AI processing [1][2][3]. Group 1: AI Inference Dilemma - The key contradiction in AI large model inference lies between the prefill and decode stages, which have opposing hardware requirements [2]. - Prefill requires high computational power but low memory bandwidth, while decode relies on high memory bandwidth with lower computational needs [3]. Group 2: Rubin CPX Configuration - Rubin CPX is designed specifically for the prefill stage, optimizing cost and performance by using GDDR7 instead of HBM, significantly reducing BOM costs to 25% of R200 while providing 60% of its computational power [4][6]. - The memory bandwidth utilization during prefill tasks is drastically improved, with Rubin CPX achieving 4.2% utilization compared to R200's 0.7% [7]. Group 3: Oberon Rack Innovations - Nvidia introduced the third-generation Oberon architecture, featuring a cable-free design that enhances reliability and space efficiency [9]. - The new rack employs a 100% liquid cooling solution to manage the increased power demands, with a power budget of 370kW [10]. Group 4: Competitive Landscape - Nvidia's advancements have intensified competition, particularly affecting AMD, Google, and AWS, as they must adapt their strategies to keep pace with Nvidia's innovations [13][14]. - The introduction of specialized chips for prefill and potential future developments in decode chips could further solidify Nvidia's market position [14]. Group 5: Future Implications - The demand for GDDR7 is expected to surge due to its use in Rubin CPX, with Samsung poised to benefit from increased orders [15][16]. - The article suggests that companies developing custom ASIC chips may face challenges in keeping up with Nvidia's rapid advancements in specialized hardware [14].
美国不准三星和SK升级扩建中国工厂,中国这个技术领域突破要抓紧
Xin Lang Cai Jing· 2025-09-05 06:25
Core Viewpoint - The U.S. government has revoked the "Verified End User (VEU)" status for Samsung Electronics and SK Hynix's factories in China, requiring them to obtain separate licenses for each batch of equipment shipped starting next year, which could lead to production delays and increased global memory prices [1][3][18]. Group 1: U.S. Regulatory Actions - The U.S. has closed exemptions for Samsung and SK Hynix, preventing them from using American machinery to produce high-end chips in China [3][18]. - The new regulations will require both companies to seek U.S. approval for each shipment to China, potentially causing significant delays and inefficiencies [3][18]. - The U.S. Department of Commerce stated that while the current situation can be maintained, no expansion of production or technological upgrades will be allowed, effectively blocking upgrades and expansions in China [20]. Group 2: Market Impact - The global semiconductor market is projected to grow by 18% in 2024, reaching $626 billion, with Samsung regaining the top position in sales due to a rebound in memory chip prices [9][12]. - Samsung's Xi'an factory produces 40% of its global NAND output, and SK Hynix is expected to surpass Samsung in sales by Q2 2025, indicating a competitive shift in the market [12][21]. - The price of SK Hynix's HBM3 memory has increased over five times since 2023, highlighting the growing demand for high-bandwidth memory in AI applications [16][21]. Group 3: Competitive Landscape - Samsung and SK Hynix dominate the HBM memory market, controlling 90% of global production, which is critical for AI chip performance [16][22]. - The U.S. is concerned that Samsung and SK Hynix could produce advanced AI memory components in China, prompting regulatory actions to slow down China's progress in this area [21][24]. - The restrictions on Samsung and SK Hynix may create opportunities for Chinese companies to develop their own high-end HBM memory, as they seek to fill the market gap left by these restrictions [24][25].
又一巨头,发力先进封装
半导体行业观察· 2025-08-16 03:38
Core Viewpoint - Samsung Electronics has signed a $16.5 billion chip foundry deal with Tesla, which boosts market confidence and offers a glimmer of hope for its long-struggling foundry business [2][5]. Group 1: Foundry Business Challenges - Samsung has faced significant challenges in the foundry sector, particularly in advanced process technology, where it initially struggled with yield issues in its 3nm process, leading to a loss of high-end customer orders to TSMC [2][5]. - Market research indicates that TSMC's global foundry market share reached 67.6% in Q1 2025, while Samsung's share dropped from 8.1% to 7.7% [2]. - Samsung has postponed the mass production of its 1.4nm process from 2027 to 2029, highlighting difficulties in expanding its advanced process market [2]. Group 2: Advanced Packaging Strategy - In response to challenges in the foundry market, Samsung is focusing on advanced packaging technology as a strategic path to breakthrough, planning to invest $7 billion in a new advanced chip packaging factory in the U.S. [3][5]. - The new packaging factory aims to address the current gap in high-end packaging technology in the U.S., where 90% of advanced packaging capacity is concentrated in Asia [5][6]. - This factory will be a key part of Samsung's integrated "design-manufacture-package" model, aiming to provide comprehensive services from chip design to product delivery [5][6]. Group 3: Market Positioning and Collaboration - The recent Tesla order significantly boosts Samsung's market confidence and supports its plans for further investment in the U.S. market [5][6]. - Samsung's strategy includes establishing local packaging facilities to meet the urgent demand for localized production, especially in light of U.S. tariffs [6]. - The advanced packaging market is projected to grow from $34.5 billion in 2023 to $80 billion by 2032, providing a strong incentive for Samsung to enhance its capabilities [9]. Group 4: Technological Innovations - Samsung is advancing its System on Panel (SoP) technology to challenge TSMC's System on Wafer (SoW) dominance, focusing on larger panel sizes for better integration of AI chips [10][11]. - The company is also investing in glass substrate technology, aiming for a 2028 rollout to replace traditional silicon substrates, which could lower costs and improve performance [16][17]. - Samsung's Fan-Out Packaging (FOPKG) technology is designed to meet the demands of mobile AI chips, achieving significant improvements in production efficiency and thermal management [19][20]. Group 5: Competitive Landscape - Samsung's advanced packaging efforts are seen as a direct challenge to TSMC's market leadership, with the company aiming to close the gap in high-end packaging capabilities [9][10]. - The establishment of a research center in Yokohama, Japan, with a $1.7 million investment, underscores Samsung's commitment to enhancing its technological prowess in advanced packaging [8]. - The competitive landscape in advanced packaging is intensifying, with Samsung's initiatives expected to reshape the global semiconductor industry [46].
全球芯片代工增长17%!
国芯网· 2025-07-28 14:03
Core Insights - The global pure semiconductor foundry industry revenue is projected to reach $165 billion by 2025, representing a 17% year-on-year growth, with a compound annual growth rate (CAGR) of 12% from 2021 to 2025, driven primarily by advanced process nodes [2] - Revenue from the 3nm node is expected to grow over 600% year-on-year, reaching $30 billion, while the 5/4nm nodes are anticipated to exceed $40 billion, contributing more than half of the total revenue from pure foundries by 2025 [2] - The demand for high-end smartphones, AI PC solutions, AI ASICs, GPUs, and high-performance computing (HPC) solutions is the main driver behind the revenue growth of advanced processes [2] Industry Competition Landscape - TSMC holds a leading position in advanced nodes, followed by Samsung and Intel, while UMC, GlobalFoundries, and SMIC continue to see strong demand in other nodes, although their revenue growth may not match that of advanced nodes [3] - Innovations in backend packaging processes, such as HBM memory integration and the shift towards chip-scale packaging, are creating new growth opportunities for the industry [3][4] - These innovations not only enhance product performance and reliability but also open up new revenue streams for semiconductor foundries [4]
人工智能,重塑了处理器格局
半导体行业观察· 2025-07-21 01:22
Core Insights - The processor market is expected to grow significantly, driven by the increasing demand for generative AI applications, with market size projected to rise from $288 billion in 2024 to $554 billion by 2030 [1] - The GPU market is anticipated to surpass the APU market for the first time in 2024, reflecting the high computational power demand, particularly in server applications [1] - The competition in the GPU market is intensifying due to the development of AI ASIC chips by major players like Google and AWS, aimed at reducing capital expenditure [1][12] - The data center processor market is rapidly expanding, projected to reach $147 billion in 2024 and $372 billion by 2030, primarily driven by generative AI applications [9] Market Dynamics - The processor market is highly concentrated, with Intel holding 66% of the CPU market and Nvidia over 90% of the GPU market, while the APU and AI ASIC & DPU markets are more fragmented [3] - New entrants in the processor market, particularly from China, are emerging, with companies like Xiaomi and NIO achieving success in specific segments [3][4] - The trend towards advanced technology nodes is evident across all segments, with a significant reduction in the number of foundries capable of producing cutting-edge nodes [7] Technological Advancements - The transition to smaller technology nodes is crucial, with CPUs expected to adopt 3nm processes by 2024, while GPUs and AI ASICs are still on 4nm processes [15] - The demand for AI applications has led to an 8-fold increase in computing performance since 2020, with Nvidia's upcoming Rubin Ultra expected to achieve 100 PetaFLOP inference speeds by 2027 [15] - The integration of HBM memory in AI solutions is critical, although several AI ASIC startups are exploring SRAM-based processors for enhanced performance [15] Strategic Developments - Governments are investing in dedicated AI data centers to ensure national computing capabilities, while the U.S. government is implementing strict export controls affecting China's access to advanced AI chips [18] - In response, China is accelerating its semiconductor industry development, with companies like Huawei focusing on CPU and AI ASIC advancements [18] - Strategic computing is becoming central to AI infrastructure, with significant investments and mergers occurring in the AI chip sector, highlighting the increasing value of silicon expertise [19]
疯狂备货,英伟达盯上了这类存储
半导体行业观察· 2025-07-17 00:50
Core Viewpoint - NVIDIA is set to produce a significant inventory of its modular memory solution, SOCAMM, to enhance the performance and efficiency of its AI products, with an expected production of 600,000 to 800,000 units this year [3][4][9]. Summary by Sections SOCAMM Memory Overview - SOCAMM memory is based on LPDDR DRAM, traditionally used in mobile devices, and is designed to be modular and upgradeable, differing from HBM and LPDDR5X solutions [3][4]. - The bandwidth of SOCAMM memory is approximately 150-250 GB/s, making it a versatile option for AI PCs and servers [5]. Production and Market Impact - The initial production target of 800,000 units is lower than the HBM memory supplied to NVIDIA by its partners, but production is expected to ramp up next year with the introduction of SOCAMM 2 [4][6]. - Major memory manufacturers, including Micron, Samsung, and SK Hynix, are competing to establish a foothold in the emerging SOCAMM market, which is anticipated to be as strategically important as the HBM market [6][10]. Technical Advantages - SOCAMM is designed for low power consumption, with its power requirements being significantly lower than traditional DDR5 modules, making it suitable for large-scale deployment [7][9]. - The modular design allows for easy upgrades, and it is expected to be used alongside HBM chips in NVIDIA's upcoming AI accelerators [7][9]. Competitive Landscape - Micron has begun mass production of SOCAMM modules, claiming a bandwidth over 2.5 times that of RDIMM and a power consumption only one-third of RDIMM [9]. - Samsung aims to regain its leadership in the DRAM market by leveraging its dominance in LPDDR technology to enter the SOCAMM market [10].