Memory Moves Into the Compute Package 存储向计算封装内部推进

Hot Chips roadmaps put DRAM closer to logic as bandwidth, power, and thermal limits tighten, while capital and policy determine where the supporting supply chain can expand. 随着带宽、功耗与散热约束收紧，Hot Chips披露的路线图正把DRAM推向更靠近逻辑的位置；与此同时，资本与政策正在决定配套供应链能够在哪里扩张。

AI & Accelerators AI与加速器

Nvidia Reportedly Earmarks $7 Billion for Poolside 报道称Nvidia为Poolside预留$7B

Wccftech · 7h ago Wccftech · 7小时前

Nvidia has reportedly earmarked $7 billion for Poolside in an arrangement that could make the chipmaker a backstop buyer for GPUs used by the AI startup. The deal also links Nvidia more closely to Poolside's open-weight model work and its own Nemotron ecosystem. The scale matters because vendor-backed financing and purchase commitments can sustain accelerator demand while shifting more credit risk onto the supplier. 报道称，Nvidia已为Poolside预留$7B，相关安排可能使这家芯片公司成为该AI初创企业所用GPU的兜底买家。交易还将Nvidia与Poolside的开放权重模型工作及其Nemotron生态绑定得更紧。其规模值得关注，因为供应商支持的融资与采购承诺既能托住加速器需求，也会把更多信用风险转移到供应商一侧。

Foundry & Manufacturing 晶圆代工与制造

NXP Adds a 500,000-Square-Foot Malaysia Assembly and Test Factory NXP在马来西亚新增50万平方英尺封测工厂

EE Times Asia · 1h ago EE Times Asia · 1小时前

NXP is expanding its Malaysian back-end operations with a 500,000-square-foot smart factory in Petaling Jaya, with production scheduled to ramp in 2028. The project adds assembly-and-test capacity rather than wafer fabrication, reinforcing Malaysia's role in the global semiconductor back end. The long lead time also shows that suppliers are planning capacity beyond the current AI-driven cycle. NXP将在马来西亚Petaling Jaya建设一座50万平方英尺的智能工厂，计划于2028年爬坡投产，以扩充后段制造能力。该项目增加的是封装与测试产能，而非晶圆制造，进一步强化马来西亚在全球半导体后端环节的地位。较长的建设周期也表明，供应商正在为当前AI周期之后的需求配置产能。

Itochu and a Taiwan IT Firm Back Chip Expansion in Southern Japan 伊藤忠与台湾IT企业支持芯片产业向日本南部扩张

Nikkei Asia · 4h ago Nikkei Asia · 4小时前

Itochu and a Taiwanese IT company are supporting the expansion of Taiwanese semiconductor businesses into southern Japan. The effort broadens the supplier base around Japan's growing chip-manufacturing footprint and lowers the practical barriers for smaller Taiwanese vendors entering the market. It is another sign that fab investment is pulling service and materials ecosystems across borders with it. 伊藤忠与一家台湾IT企业正在支持台湾半导体公司向日本南部扩张。该行动将扩充日本不断增长的芯片制造版图周边供应商基础，并降低台湾中小厂商进入当地市场的实际门槛。这再次表明，晶圆厂投资正在带动服务与材料生态同步跨境迁移。

Memory 存储

SK hynix Maps EMIB and 3D Packaging Into Its HBM Roadmap SK hynix将EMIB与3D封装纳入HBM路线图

Wccftech · 4h ago Wccftech · 4小时前

SK hynix used Hot Chips 2026 to outline advanced-packaging options for future HBM, including Intel's EMIB and an eventual move toward 3D structures. The roadmap treats packaging as a primary scaling lever as conventional HBM stacks encounter interconnect, thermal, and package-area constraints. Execution will depend on yields and ecosystem support across logic, memory, and interposers. SK hynix在Hot Chips 2026上披露了未来HBM的先进封装选项，包括Intel EMIB，并计划最终迈向3D结构。随着传统HBM堆叠面临互连、散热和封装面积约束，该路线图把封装视为主要扩展杠杆。其落地仍取决于逻辑、存储和中介层生态的良率与协同。

Samsung's zHBM Roadmap Stacks DRAM on Compute Samsung zHBM路线图将DRAM堆叠于计算之上

Wccftech · 6h ago Wccftech · 6小时前

Samsung presented a three-phase HBM roadmap that gives the base die more logic functions and culminates in zHBM, with DRAM stacked directly on compute. Moving logic into the memory stack could reduce data movement and reclaim package area, but it also tightens thermal and process-integration requirements. The disclosure positions HBM base-die design, not only DRAM speed, as a competitive axis. Samsung公布了一条三阶段HBM路线图：逐步为base die加入更多逻辑功能，最终迈向把DRAM直接堆叠在计算芯片上的zHBM。将逻辑移入存储堆叠有望减少数据搬运并释放封装面积，但也会提高散热与制程集成要求。这一披露说明，HBM竞争正从DRAM速度延伸到base die设计。

Micron Says Compute Growth Is Outrunning HBM Bandwidth Micron称计算增长速度正超过HBM带宽

Wccftech · 7h ago Wccftech · 7小时前

Micron says AI compute capability is outpacing HBM bandwidth by roughly 3x every two years, worsening the memory wall. Its Hot Chips discussion also cited HBM failures as the cause of 17% of interruptions during Meta's Llama 3 training. The figures make bandwidth, reliability, packaging, and heat removal joint constraints rather than separable memory specifications. Micron称，AI计算能力相对HBM带宽的增长速度约为每两年拉开3倍，内存墙因此继续恶化。其Hot Chips演讲还指出，Meta训练Llama 3期间有17%的中断由HBM故障引起。这些数据表明，带宽、可靠性、封装与散热已经成为相互耦合的约束，而非彼此独立的存储规格。

Equipment & Materials 设备与材料

Underlayer Engineering Targets Lower EUV Exposure Dose 底层材料工程瞄准更低EUV曝光剂量

SemiWiki · 4h ago SemiWiki · 4小时前

New analysis highlights resist underlayer engineering as a route to reduce EUV exposure dose without surrendering resolution. The materials problem remains a four-way trade-off among dose, line-edge or linewidth roughness, resolution, and defectivity. Lower dose can raise scanner throughput, but only if the full resist stack preserves pattern fidelity at production scale. 最新分析指出，光刻胶底层材料工程可在不牺牲分辨率的前提下降低EUV曝光剂量。材料体系仍需在剂量、线边缘或线宽粗糙度、分辨率与缺陷率之间进行四重权衡。降低剂量可以提升光刻机吞吐量，但前提是完整光刻胶堆叠在量产尺度上保持图形保真度。

Policy & Geopolitics 政策与地缘政治

Samsung and SK hynix Keep Investing in China Fabs Despite U.S. Curbs 尽管美国限制加码，Samsung与SK hynix仍投资中国工厂

Korea JoongAng Daily · 5h ago Korea JoongAng Daily · 5小时前

Samsung and SK hynix are directing fresh investment into aging Chinese fabs despite U.S. restrictions on advanced semiconductor equipment. The spending seeks to preserve the competitiveness and useful life of large installed manufacturing bases without assuming unrestricted access to leading-edge tools. That balancing act makes licensing policy a direct variable in memory supply and capital allocation. 尽管美国限制先进半导体设备流入中国，Samsung与SK hynix仍在向其老化的中国工厂投入新增资本。相关支出旨在不依赖无限制获取尖端设备的前提下，维持既有大规模制造基地的竞争力并延长使用寿命。这种平衡使许可政策直接成为存储供应与资本配置的变量。

Earnings & Markets 业绩与市场

AI-Chip Demand Lifts Korean Materials Makers' Second-Quarter Profits AI芯片需求推升韩国材料厂商二季度利润

DIGITIMES · 1h ago DIGITIMES · 1小时前

South Korean semiconductor-materials suppliers reported stronger profits in the second quarter of 2026 as AI-chip production raised demand upstream. The gains show that value from the accelerator buildout is reaching specialty-material vendors, not only GPU and memory manufacturers. The next test is whether orders remain broad as customers concentrate spending on the most constrained AI components. 随着AI芯片生产带动上游需求，韩国半导体材料供应商在2026年第二季度录得更强利润。这表明加速器扩产创造的价值正在传导至特种材料厂商，而不仅限于GPU与存储制造商。下一步要观察的是，当客户把支出集中到最紧缺的AI部件时，材料订单能否继续保持广度。

Challengers & New Silicon 新势力与新硅

d-Matrix Puts 3D DRAM Beneath Compute in Raptor d-Matrix Raptor将3D DRAM置于计算层下方

Wccftech · 3h ago Wccftech · 3小时前

d-Matrix disclosed Raptor, a 3D-DRAM architecture that removes the conventional memory PHY and stacks DRAM beneath compute for generative-AI inference. The company claims SRAM-class bandwidth at one-tenth the power of HBM, directly targeting the cost of moving weights and KV-cache data. The Hot Chips disclosure is an architectural claim rather than an independent production benchmark, but it offers a concrete alternative to scaling external HBM stacks. d-Matrix披露了Raptor 3D-DRAM架构：取消传统存储PHY，并把DRAM堆叠在计算层下方，用于生成式AI推理。公司宣称其带宽可达到SRAM级别，而功耗仅为HBM的十分之一，直接瞄准权重与KV cache数据搬运成本。Hot Chips披露目前仍是架构层面的主张，而非独立量产基准，但它为继续扩展外置HBM堆叠提供了具体替代路径。