🔥36氪•較早收集於 8m
三星1c DRAM良率超80%,HBM4逼近60%
#memory-yield#hbm-production#dram-testingsamsung-1c-dram-/-hbm4samsung1c-dramhbm4
💡Samsung HBM4 yield jump eases AI GPU memory shortages for training/inference
⚡ 30-Second TL;DR
有什麼變化
1c DRAM高溫熱測試良率超80%
為什麼重要
提升三星在AI GPU用HBM競爭力,可能穩定Nvidia H100/H200及Blackwell晶片記憶體供應。
下一步行動
Benchmark Samsung HBM4 prototypes against SK Hynix for your next-gen AI cluster memory selection.
誰應關注:Enterprise & Security Teams
關鍵要點
- •1c DRAM高溫熱測試良率超80%
- •預計2025年5月達90%良率
- •HBM4良率升至近60%,較2024年Q4的50%提升
- •三星高溫環境下取得最高良率
🧠 深度解析
背景與延伸:來自公開資料,非原文內容。引用 4 個來源。
🔑 增強重點摘要
- •Samsung redesigned its 1c DRAM process under new leadership by increasing peripheral circuit line width to prioritize stability over cost, addressing prior lag behind SK hynix and Micron who used advanced 1b DRAM for HBM3E.[1]
- •Samsung pursued dual strategies for 1c DRAM improvement: modifying prior 1a/1b designs or full redesign with larger chips for higher yields, though at increased cost due to more wafer usage.[1][2]
- •HBM4 logic die, fabricated on Samsung's 4nm foundry process, achieved over 80% mass-production yield, exceeding 90% in early tests before performance tweaks, enabling stable commercialization.[3]
- •NVIDIA requested Samsung to boost 1c DRAM yields for reliable HBM4 supply, as 1c will extend to HBM4E, prompting design changes amid competitors' mature 1b DRAM processes over 80% yield.[3]
📊 競品分析▸ Show
| Aspect | Samsung (1c DRAM for HBM4) | SK hynix/Micron (1b DRAM for HBM3E) |
|---|---|---|
| DRAM Generation | 6th gen (1c, newer) | 5th gen (1b, 10nm class) |
| Yield Status | 1c ~60%, logic die >80% | >80%, mass prod 1-2 years stable |
| HBM Market Pos. | Catching up, redesigning | Leading, established supply |
🛠️ 技術深入
- •1c DRAM is sixth-generation 10nm-class DRAM, one generation ahead of competitors' 1b (5th-gen 10nm) used in HBM3E core dies.
- •Redesign increases peripheral circuit line width for chip stability; full redesign option yields larger chips with higher production rates but elevated costs from increased wafer consumption.
- •HBM4 logic die uses Samsung Foundry's 4nm process as the 'brain' for stacking and control.
🔮 前景展望AI analysis grounded in cited sources
Samsung HBM4 samples to NVIDIA in Q3 2025
Improved 1c yields enable sampling to NVIDIA for next-gen AI chips in late 2025 per reports on redesign progress.[1]
HBM4 mass production delayed to 2026
1c yield challenges, despite progress to 65% in limited runs, push full rollout beyond 2025 as confirmed by industry sources.[2]
1c DRAM yields to hit mass production viability by H2 2025
Q2 2025 yields at 30-40% with rapid improvement signal full-scale production start in second half of 2025.[1]
⏳ 時間線
2024-Q4
1c DRAM yields at 60-70%; HBM4 at 50%
2025-06
1c DRAM yields reach 50-70%; redesign initiated under new chairman
2025-Q2
1c DRAM yields improve to 30-40%
2025-07
Limited 1c wafer runs hit ~65% yield; HBM4 rollout delayed to 2026
2026-02
1c DRAM exceeds 80% in high-temp tests; HBM4 nears 60% yield
📎 來源 (4)
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: 36氪 ↗
每週 AI 簡報
每週一封,可隨時退訂。