來源IT之家•較早收集於 83m
Nvidia CEO 黃仁勳將與三星電子高層會面
💡關鍵供應鏈更新：Nvidia 對 HBM 記憶體的布局將直接影響未來 AI 硬體的產能與供應。
⚡ 30 秒速覽
有什麼變化
討論重點將聚焦於 AI 基礎設施所需的高頻寬記憶體 (HBM) 及次世代記憶體技術。
為什麼重要
加強與三星在 HBM 領域的合作，對於 Nvidia 在 AI 晶片市場維持領先地位至關重要，特別是在 GPU 記憶體需求激增的情況下。
下一步行動
追蹤 Nvidia 與三星的供應鏈公告，因為 HBM 的供應狀況直接影響次世代 AI 叢集的部署時程。
誰應關注：Founders & Product Leaders
關鍵要點
- •討論重點將聚焦於 AI 基礎設施所需的高頻寬記憶體 (HBM) 及次世代記憶體技術。
- •黃仁勳正與包括 LG、SK 及現代汽車在內的韓國科技巨頭進行一系列高層會談。
- •此舉旨在強化 Nvidia 在 AI 硬體與機器人領域的供應鏈布局。
🧠 深度解析
背景與延伸：來自公開資料，非原文內容。引用 24 個來源。
🔑 增強重點摘要
- •Jensen Huang's visit is part of a broader strategy to build an "Nvidia-centered value chain" in South Korea, encompassing memory chips, robotics hardware, and gaming software, beyond just HBM supply.
- •Nvidia has certified HBM4 products from all three major memory manufacturers—Samsung, SK Hynix, and Micron—for its upcoming Vera Rubin AI accelerator platform, indicating a diversified supply strategy.
- •Samsung is accelerating its HBM development cycle from approximately two years to one year to align with the rapid annual rollout of new AI accelerators from major clients like Nvidia.
- •Nvidia is expanding its robotics partnerships beyond China to include the U.S., Europe, and South Korea, aiming to create a global robotics ecosystem powered by its AI infrastructure.
- •LG Group is investing in 10,000 Nvidia Blackwell GPUs and expanding its partnership with Nvidia into "physical AI," including robotics and smart factories, and co-developing domain-specific AI models.
📊 競品分析▸ Show
| Feature/Company | SK Hynix | Samsung | Micron |
|---|---|---|---|
| HBM Market Share (Q2 2025) | ~62% | ~17% | ~21% |
| HBM Generations | HBM3E, HBM4, iHBM (thermal design) | HBM3E, HBM4, HBM4E, HBM5 (mockup) | HBM3E, HBM4 |
| HBM4 Certification for Nvidia Rubin | Certified | Certified | Certified |
| HBM4 Bandwidth (per stack) | Up to 3.3 TB/s (HBM4) | Up to 3.6 TB/s (HBM4E) | Exceeding 2.8 TB/s (HBM4) |
| HBM4 Interface Width | 2048-bit | 2048-bit | 2048-bit |
| HBM4 Capacity (per stack) | Up to 64GB (16-Hi stacks) | Up to 64GB (16-layer configurations) | 36GB (12-Hi HBM3E), HBM4 higher |
| HBM4 Power Efficiency | Improved over HBM3E | Improved over HBM4 (HBM4E) | Superior power-efficiency specs |
| HBM Supply Status (2026) | Entire 2026 capacity sold out | Plans 50% capacity surge in 2026, delivering HBM4 samples to Nvidia | Entire 2026 capacity sold out |
🛠️ 技術深入
- HBM4 Architecture: HBM4 represents a fundamental architectural shift from HBM3e, doubling the memory interface to 2048-bit and integrating a logic base die.
- Bandwidth: HBM4 targets over 2.0 TB/s per stack, with advanced configurations reaching up to 3.3 TB/s, a significant increase from HBM3e's 1.2 TB/s.
- Capacity: HBM4 supports up to 64GB per stack through 16-Hi stacks and 32Gb layers, enabling larger models.
- Power Efficiency: HBM4 aims for lower operating voltage (0.9-1.0V vs. 1.1V in HBM3e) and improved efficiency, targeting ~120 GB/s/W.
- Interposer Redesign: The wider 2048-bit interface of HBM4 requires next-generation interposer technologies with sub-micron routing features to manage increased signal density and prevent crosstalk.
- Logic Base Die: The HBM4 base die is built on a logic process (e.g., 12nm/5nm), requiring the host GPU or ASIC controller to be updated to leverage its capabilities.
- HBM4E: Samsung's HBM4E samples deliver a stable pin speed of 14 Gbps, scalable up to 16 Gbps, providing memory bandwidth of up to 3.6 TB/s per stack.
- Cooling Innovations: Both Samsung (Heat Path Block - HPB) and SK Hynix (iHBM) are developing in-package cooling structures to address thermal bottlenecks in the die-to-die interface of HBM stacks.
🔮 前景展望基於引用來源的 AI 分析
The HBM supply shortage will persist through 2026 and potentially until 2030.
The demand for HBM from the AI sector is rapidly increasing, with next-generation AI accelerators requiring significantly more memory, and manufacturing capacity cannot scale fast enough to meet this demand.
Nvidia's diversified HBM supplier strategy will reduce supply chain risks for its next-generation AI accelerators.
By certifying HBM4 from all three major manufacturers (SK Hynix, Samsung, Micron), Nvidia ensures greater flexibility, reduces dependence on a single supplier, and prepares for the next wave of AI infrastructure investment.
South Korea is becoming a central hub for Nvidia's "physical AI" ecosystem development.
Jensen Huang's extensive meetings with major Korean conglomerates (Samsung, SK, LG, Hyundai) across memory, robotics, and automotive sectors indicate a strategic focus on leveraging Korean expertise for AI and robotics.
⏳ 時間線
2015
Hyundai Motor Group begins partnership with NVIDIA for intelligent systems.
2024-07
JEDEC announces preliminary specifications for HBM4.
2025-10
NVIDIA and SK Group announce plans to build an AI factory in South Korea and deepen collaboration on HBM.
2026-03
NVIDIA expands partnerships with global robotics leaders for physical AI development.
2026-04
LG Group and NVIDIA agree to expand technical cooperation for next-generation, domain-specific AI models and ecosystems.
2026-06
Samsung begins shipping industry-first 12-layer HBM4E samples to major global customers.
2026-06
NVIDIA certifies HBM4 from SK Hynix, Samsung, and Micron for its upcoming Vera Rubin platform.
📎 來源 (24)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： IT之家 ↗
每週電子報
每週一封，可隨時退訂。