AI 加速器晶片常受三種供應短缺影響:HBM 記憶體、CoWoS 封裝產能,以及台積電 3nm 製程的產能。
- Cerebras 共同創辦人暨執行長 Andrew Feldman 說,Cerebras 的設計避開了前兩項瓶頸,目前晶片則採用 5nm 製程,沒有依賴 3nm 製程。
- 由於記憶體、先進封裝和尖端製程的供應都可能限制晶片出貨,Cerebras 因此有不同於主流 GPU 系統的供應條件。
AI 伴隨問答
針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:
Andrew Feldman, co-founder and CEO of Cerebras explains the 3 major supply-chain bottlenecks affecting much of the AI accelerator industry:
HBM memory, CoWoS packaging capacity, and access to TSMC’s 3nm manufacturing. Cerebras largely sidesteps all 3 by architectural choice.
It does not rely on HBM or CoWoS, and its current chips are manufactured on a 5nm process rather than 3nm.
In an industry where access to memory, advanced packaging, and leading-edge fabrication capacity can directly constrain shipments, that gives Cerebras a materially different supply profile from the main-street GPU systems.
----
full video on The MAD Podcast with Matt Turck and Cerebras YouTube channel, (link in comment)