加速卡Accelerators
Cerebras WSE-3T(晶圆级)Cerebras WSE-3T (wafer-scale)
回答「晶圆级芯片单片的晶体管、面积、核心数与官方 AI 算力口径」。250 PF 是官方“AI compute”表述,未注明是 dense FP16 还是特定精度,不能直接和 GPU 的 FP16/FP8 数字横比。
不主张:不是单卡 PCIe 加速器;不是 CS-4 整机算力;未公开对外售价与功耗口径。Answers the transistor count, die area, core count and the vendor's "AI compute" figure for one wafer-scale chip. 250 PF is the vendor's AI-compute wording, not labelled as dense FP16 or a specific precision, so it is not directly comparable with GPU FP16/FP8 numbers.
It does not claim: this is not a PCIe add-in card; it is not the CS-4 system figure; price and power are not stated here.
关键参数Key specs
- 晶体管
- 4 万亿
- 芯片面积
- 46,225 mm²
- AI 核心
- 900,000
- 官方 AI 算力
- 250 petaflops(口径见原文)
- 系统形态
- CS-4 机架系统(3 颗 WSE-3T)
官方原文怎么说What the source says
Four trillion transistors. 250 petaflops. One silicon wafer.
46,225 mm² and containing 4 trillion transistors
900,000 AI-optimized cores
The Cerebras CS-4 rack-scale AI system is powered by three WSE-3T processors
资料与许可Source and licence
- 资料时间Source checked
- 2026-11-19
- 许可 / 类型Licence / kind
- official-spec-page
- 许可说明Licence evidence
- https://www.cerebras.ai/product-chip/
- 类目Category
- 加速卡Accelerators