semi·news
Headlines要闻 / Research研究 /
Tuesday, May 26, 2026 2026年5月26日 星期二

Memory's Heat Wall Cracks Open as Taiwan's AI Rally Eclipses India 存储热墙松动,台湾AI行情盖过印度

SK hynix moves cooling inside the HBM stack just as TSMC's surge lifts Taiwan past India to the world's fifth-largest equity market — and Broadcom, AMD and Google push their custom-silicon offensive to challenge Nvidia's grip. SK海力士把散热结构嵌入HBM内部之际,台积电拉动台湾股市规模超越印度跃居全球第五;同时博通、AMD与谷歌以定制硅片合围英伟达。

Memory 存储

SK hynix puts cooling inside the HBM stack with 'iHBM' — 30% lower thermal resistance, targeted at HBM5 SK海力士将散热结构嵌入HBM内部:'iHBM'热阻降低30%,瞄准HBM5

Tom's Hardware · 2h ago Tom's Hardware · 2小时前

SK hynix unveiled iHBM, embedding silicon-based Integrated Cooling Elements (ICEs) directly into the D2D PHY region between HBM and the GPU — where heat is most concentrated — using its Mass Reflow Molded Underfill (MR-MUF) wafer-level packaging. The company quotes a 30% reduction in thermal resistance versus conventional HBM and slates the technology for next-generation parts including HBM5. The move recasts the memory wall as a thermal wall: with HBM3E/HBM4 stacks already throttling GPU performance at high duty cycles, integrated cooling becomes the gating supplier capability rather than just stack height or I/O speed. SK海力士发布iHBM技术,将硅基集成冷却元件(ICE)直接嵌入HBM与GPU之间热量最集中的D2D PHY区域,并通过其大量回流模塑底部填充(MR-MUF)晶圆级封装实现量产。官方数据称热阻较常规HBM下降30%,并将率先用于下一代HBM5。这把'存储墙'重新定义为'热墙':当HBM3E/HBM4在高负载下已开始限制GPU性能时,集成散热将成为新的关键供应商能力,超越堆叠层数或I/O速率本身。

Samsung's P4 HBM expansion risks deepening the 2027 DRAM crunch 三星P4厂HBM扩产或加剧2027年DRAM紧缺

Digitimes (via Google News) · 9h ago Digitimes (via Google News) · 9小时前

Samsung is converting more capacity at its P4 line toward HBM4 and HBM4E, but the conversion sucks wafer starts away from commodity DDR5 just as PC and server demand recovers. Analysts cited by Digitimes now expect the DRAM shortage that began this year to extend through 2027, with non-HBM contract prices likely to keep climbing through 2026. The dynamic explains why hyperscalers are signing multi-year LTAs at premium pricing — supply elasticity has effectively been mortgaged to the HBM build-out. 三星正将P4产线更多产能转向HBM4/HBM4E,但此举从通用DDR5的晶圆投片中抽走资源,恰逢PC和服务器需求回暖。Digitimes引述分析师预计:今年开始的DRAM紧缺将延续至2027年,非HBM的合约价在2026年内或继续上行。这也解释了为何超大规模客户愿以溢价签下多年LTA——产能弹性已被HBM扩产抵押。

AI & Accelerators AI与加速器

AMD, Broadcom and Google sharpen the anti-Nvidia offensive as custom silicon gains traction AMD、博通与谷歌强化'反英伟达'攻势,定制硅片势头扩大

The Economy · 1h ago The Economy · 1小时前

Broadcom's Q1 FY26 AI revenue hit $8.4 billion (+106% YoY) and the company guided Q2 AI semiconductor revenue to $10.7 billion — the clearest sign yet that custom XPU work co-designed with Google, Meta and OpenAI is no longer a sideshow to Nvidia GPUs. AMD is positioning its MI450 + Helios rack-scale platform for a 2H26 launch, while Broadcom's Tomahawk 6 102.4 Tbps Ethernet switch entered volume production in March alongside the Jericho 4 fabric. The market narrative is shifting from 'GPU vs. nothing' to a portfolio fight where general-purpose GPUs and purpose-built ASICs each take a defined share of the AI-infrastructure dollar. 博通财年2026 Q1的AI业务收入达到84亿美元(同比+106%),并指引Q2 AI半导体收入冲上107亿美元,这是与谷歌、Meta、OpenAI联合定制XPU路线'成势'的最清晰信号。AMD将MI450 + Helios机架级平台定档在2026年下半年发布;与此同时博通的Tomahawk 6 (102.4 Tbps以太网) 已于今年3月量产,搭配Jericho 4扩展织物。市场的叙事正从'英伟达GPU独大'转向'通用GPU与专用ASIC各取一份AI基础设施钱包'的格局战。

GUC showcases VSORA's Jotunn8 datacenter inference processor at TSMC Europe symposium GUC在TSMC欧洲技术论坛展示VSORA数据中心推理芯片Jotunn8

EE Times · 50m ago EE Times · 50分钟前

Taiwan's Global Unichip (TSMC affiliate, ticker 3443.TW) delivered turnkey ASIC services from netlist to manufacturing for VSORA's Jotunn8 — a European-designed inference accelerator built on TSMC 5nm with CoWoS packaging, explicitly targeting the memory-wall bottleneck for LLM serving. The collaboration matters less for its raw specs and more for what it signals: GUC's design-services model is increasingly the route by which non-US fabless startups reach leading-edge silicon, and Europe now has a credible data-center inference contender flowing through TSMC's queue. 台积电关联公司创意电子(GUC,3443.TW)以一站式ASIC服务(从网表到量产)协助VSORA完成Jotunn8 — 一款面向LLM推理'内存墙'问题的欧洲设计加速器,采用台积电5nm工艺与CoWoS封装。这一合作意义不在性能数字,而在信号本身:GUC的设计服务模式正成为非美系无晶圆厂初创进入先进工艺的主要通道,而欧洲也由此拥有了一家走台积电产能的数据中心推理候选者。

Foundry & Manufacturing 晶圆代工与制造

Taiwan overtakes India as the world's fifth-largest stock market on TSMC's AI surge 台积电拉动台湾股市市值反超印度,跃居全球第五

Bloomberg (via Google News) · 4h ago Bloomberg (via Google News) · 4小时前

Taiwan's market capitalisation moved ahead of India this week — a milestone almost wholly attributable to TSMC, whose share price has roughly tracked the AI-inference build-out and whose weighting now defines the TAIEX. With Korea's KOSPI up ~80% year-to-date and TAIEX up ~40%, the two Asian fabs-anchored markets have been the major beneficiaries of the hyperscale capex cycle that began in 2024. The flip side: passive-flow concentration into TSMC and SK hynix/Samsung gives a small basket of chipmakers outsized control of Asia-Pacific index performance. 本周台湾股市总市值反超印度,这一里程碑几乎完全来自台积电的拉动 — 其股价大体跟随AI推理基础设施的扩张,且权重已主导加权指数。年内韩国KOSPI涨约80%、台湾加权指数涨约40%,两地以晶圆厂为锚的市场成为2024年开启的超大规模资本开支周期的最大受益者。代价是:被动资金过度集中于台积电与SK海力士/三星,使一小批芯片厂的表现实质决定了亚太指数的走势。

TSMC promises faster 2026 bonus growth after employees float a Samsung-style strike TSMC员工酝酿'三星式'罢工后承诺加快2026年奖金增长

Tom's Hardware · 2h ago Tom's Hardware · 2小时前

Reports of a rumored ~15% cut to 2025 employee bonuses despite record profits sparked widespread anger among TSMC staff this week, with internal forums openly discussing unionization and a Samsung-style work stoppage. Management responded by promising materially faster bonus growth in 2026. The episode is a rare labor flashpoint for a company that has historically dominated Taiwan's talent market — and a reminder that TSMC's ramp of Arizona, Kumamoto and Dresden fabs has been depending on stretched senior engineering capacity at home. 尽管利润创纪录,关于2025年员工奖金被砍约15%的传闻引爆了台积电内部不满,公司论坛公开讨论组织工会和效仿三星式停工。管理层随即承诺2026年奖金增速将明显加快。对于历来主导台湾人才市场的台积电而言,这是一次罕见的劳资矛盾爆点 — 也提醒外界:亚利桑那、熊本、德累斯顿三地新厂的爬坡,一直建立在台湾本部资深工程师人力被拉满的基础上。

Policy & Geopolitics 政策与地缘

After $2.5B Supermicro smuggling bust, Nvidia CEO publicly urges OEMs to tighten compliance Supermicro 25亿美元走私案后,黄仁勋公开敦促OEM收紧出口合规

TechTimes · 1h ago TechTimes · 1小时前

Following Taiwan's first criminal prosecution tied to AI-chip smuggling — a case implicating $2.5 billion in alleged diverted Nvidia GPUs routed through Supermicro channels — Jensen Huang publicly called on system OEMs to overhaul their export-control processes. The episode lays bare a structural awkwardness: Nvidia's commercial relationships with server OEMs run far ahead of the visibility either party has into end-customer destinations, and the BIS now expects upstream chip vendors to police downstream channel behavior they don't actually control. 在台湾首例与AI芯片走私相关的刑事起诉之后 — 该案涉嫌通过Supermicro渠道转移约25亿美元的英伟达GPU — 黄仁勋公开呼吁整机OEM重做出口合规流程。这暴露出结构性尴尬:英伟达与服务器OEM的商业关系,远远跑在双方对最终用户去向的可见性之前;而美国商务部工业与安全局(BIS)如今期望上游芯片厂去管控自己实际并不控制的下游分销行为。

Nvidia's $200B CPU market forecast quietly includes China — even with sales still frozen 英伟达2000亿美元CPU市场预测悄然涵盖中国 — 尽管销售仍冻结

CNBC · 4h ago CNBC · 4小时前

Nvidia's investor-day projection of a $200B addressable market for data-center CPUs explicitly assumes some level of China shipment access — a quiet bet that the current US licensing freeze on H200 deliveries to Alibaba, Tencent, ByteDance and JD will lift before mid-decade. The disclosure puts the company's public narrative (Huang's 'we've largely conceded China to Huawei') in tension with its modeling. If China remains gated through 2027, the $200B figure leaves a substantial visible gap between guidance and addressable reality. 英伟达投资者日给出的数据中心CPU TAM达2000亿美元,明确假设具备一定的中国出货通道 — 这是一个安静的押注:美国对阿里、腾讯、字节、京东等公司H200交付的当前许可冻结,将在2020年代中前期解封。这与黄仁勋公开'中国市场已基本让给华为'的叙事形成张力。若中国市场到2027年仍被卡,2000亿美元指引与可达市场的差距将相当显眼。

Equipment & Materials 设备与材料

ASML to add 1,000 staff in Taiwan as service and install-base demand outruns hiring plans ASML将在台湾增聘1,000人 — 服务与装机量需求超出原招聘计划

Taiwan News (via Google News) · 10h ago Taiwan News (via Google News) · 10小时前

ASML confirmed plans to hire roughly 1,000 additional staff in Taiwan, expanding both EUV/DUV field-service capacity and the engineering footprint adjacent to TSMC's Hsinchu and Tainan fabs. The build-out reflects how install-base support — not just new tool sales — now dominates ASML's revenue trajectory: each High-NA system shipped pulls a multi-year service tail, and TSMC's parallel A16/A14 rollouts require ASML engineers literally on the fab floor. ASML确认计划在台湾再招约1,000名员工,扩充EUV/DUV现场服务团队,以及部署在台积电新竹、台南厂区附近的工程师阵地。这一扩张折射出一个趋势:装机基础维护(而非新机出货)正成为ASML收入轨迹的主导项 — 每一台High-NA系统出货都会拖出多年期的服务尾,而台积电同步推进的A16/A14节点更需要ASML工程师扎根在产线现场。

Samsung's silicon-photonics push adds a new layer to Korea-Netherlands chip ties 三星硅光子布局为韩荷半导体关系增添新层次

Digitimes (via Google News) · 15h ago Digitimes (via Google News) · 15小时前

Samsung's expansion of internal silicon-photonics R&D and its deepening collaboration with ASML and Dutch optics suppliers extends the Korea-Netherlands relationship beyond EUV lithography into co-packaged optics for AI accelerators and HBM-adjacent interconnect. The angle matters because silicon photonics is shifting from a TSMC-led story (the COUPE platform) toward a multi-foundry race — and Samsung needs the optical engine layer to keep HBM5 and successor stacks viable as I/O bandwidth scales past what copper SerDes can deliver. 三星在内部硅光子研发上的扩张,以及与ASML及荷兰光学供应商加深的合作,使韩荷半导体关系超出EUV光刻范畴,延伸进AI加速器与紧邻HBM的互连用共封装光学领域。意义在于:硅光子赛道正从台积电主导(COUPE平台)转向多家代工厂共同争夺;而三星必须在光引擎层占位,才能让HBM5及后续堆叠在I/O带宽超出铜SerDes极限后仍具可行性。

Packaging & Advanced Nodes 封装与先进节点

Advanced packaging shifts to square panels — TSMC's CoPoS pilot moves toward 310×310 mm 先进封装转向方形面板 — 台积电CoPoS试产线推进至310×310 mm

Chosun Ilbo (via Google News) · 4h ago Chosun Ilbo (via Google News) · 4小时前

TSMC's CoPoS (Chip-on-Panel-on-Substrate) pilot line is on track for June completion at the 310×310 mm panel format, with broader ramp targeted for 2028-2029. Samsung is pursuing the same panel-level packaging vector but staying on plastic substrates, while TSMC is evaluating glass — a divergence with material consequences for warpage control and signal integrity at sub-2nm chiplet partitions. Multiple panel sizes (310×310, 510×515, 600×600 mm) are now being standardized across customers including TSMC, Intel and Samsung. 台积电CoPoS(Chip-on-Panel-on-Substrate)试产线按计划于6月完成、面板规格为310×310 mm,规模化爬坡指向2028-2029年。三星走相同的面板级封装路线但仍采用塑胶基板,台积电则在评估玻璃基板 — 这一分歧对2nm以下chiplet分区的翘曲控制与信号完整性有实质性影响。310×310、510×515、600×600 mm等多种面板尺寸正在台积电、英特尔、三星等客户之间形成标准化。

Lightmatter joins TSMC's COUPE program for 3D optical engines Lightmatter加入台积电COUPE项目,共推3D光引擎

Digitimes (via Google News) · 12h ago Digitimes (via Google News) · 12小时前

Lightmatter, the photonic-compute startup, is now a named partner on TSMC's COUPE (Compact Universal Photonic Engine) co-packaged-optics platform — moving from architectural advocacy into a concrete production path for its 3D-stacked optical engine. The deal validates TSMC's bet that the next jump in switch and accelerator bandwidth (above 1.6 Tbps per port, multi-million-XPU fabrics) will be optical-first; it also tightens the pipeline between TSMC, Nvidia's Quantum-X switches, and Broadcom's Tomahawk roadmap. 光子计算初创Lightmatter成为台积电COUPE(Compact Universal Photonic Engine)共封装光学平台的明确合作伙伴 — 从架构倡导走入面向其3D堆叠光引擎的具体量产路径。该合作验证了台积电的判断:交换芯片与加速器下一代带宽(单端口1.6 Tbps以上、百万级XPU织物)将以光为主;同时也收紧了台积电、英伟达Quantum-X交换机与博通Tomahawk路线之间的衔接。