晶圆级芯片加速升温：Cerebras拿下OpenAI超200亿美元大单，特斯拉重启Dojo，中国取得多项技术突破
英文摘要
Cerebras has secured a multi-year computing power procurement agreement with OpenAI, with a potential total value exceeding $20 billion, to be delivered by 2028; the company's current wafer-scale engine WSE-3 uses a 5nm process, covers 46,225 mm², integrates 4 trillion transistors and 900,000 AI cores, peaking at 125 PFLOPS. Tesla restarted Dojo 3, relying on TSMC's InFO_SoW wafer-level packaging to integrate 25 D1 chips into a training tile. In China, a Tsinghua University team demonstrated a 12-inch wafer-scale AI chip prototype and published three ISCA papers showing up to a 3.12x average performance improvement over GPU clusters for large-model inference mapping. The Institute of Computing Technology, CAS, developed a 16-module prototype called 'Ying Tian Lake' with 1.45x and 1.78x improvements in linear algebra and inference tasks, and proposed the Ouroboros wafer-scale SRAM compute-in-memory architecture that integrates 54 GB of SRAM on a single wafer, delivering 150,000 tokens/s for Llama 13B inference. New Ziguang Group announced the 'Zixian' 3D near-memory computing architecture achieving 30 TB/s memory bandwidth and planning wafer-level integration. Wafer-level chips are moving from technical demonstrations into industrial validation.
中文摘要
Cerebras与OpenAI签订了一份潜在总价值超200亿美元、至2028年交付的算力采购协议，其当前晶圆级引擎WSE-3采用5nm工艺，面积46,225平方毫米，集成4万亿晶体管和90万个AI核心，峰值算力125 PFLOPS。特斯拉重启Dojo 3，通过台积电InFO_SoW晶圆级封装将25颗D1芯片集成为一个训练瓦片。国内，清华大学团队成功研制12英寸晶圆级AI芯片验证样机，并在ISCA发表三项成果，其中大模型推理映射方案相较GPU集群平均性能提升3.12倍。中科院计算技术研究所研发了由16个计算模组组成的“映天湖”原型，线代与推理任务分别改善1.45倍和1.78倍，并提出Ouroboros晶圆级SRAM存算一体架构，在单晶圆上集成54GB SRAM，运行Llama 13B时吞吐达15万tokens/s。新紫光集团发布“紫弦”三维近存计算架构，存储带宽30TB/s，并规划迈向晶圆级集成。晶圆级芯片正从技术验证进入产业验证阶段。
关键要点
Cerebras signed a >$20B deal with OpenAI, and its WSE-3 wafer-scale chip integrates 4 trillion transistors with 125 PFLOPS; the company is now publicly listed.
Cerebras与OpenAI签订超200亿美元大单，WSE-3晶圆级芯片集成4万亿晶体管、125 PFLOPS算力，公司已上市。
Tesla restarted Dojo 3, using TSMC InFO_SoW to integrate 25 D1 chiplets for a wafer-scale training tile.
特斯拉重启Dojo 3，利用台积电InFO_SoW集成25颗D1芯粒，构建晶圆级训练瓦片。
Tsinghua University produced a 12-inch wafer-scale AI chip prototype and its ISCA papers reported up to 3.12x performance gains against GPU clusters for inference mapping.
清华大学推出12英寸晶圆级AI芯片样机，ISCA论文显示推理映射较GPU集群最高提升3.12倍。
CAS 'Ying Tian Lake' prototype integrated 16 modules achieving 1.45x and 1.78x improvements; Ouroboros architecture packed 54 GB SRAM on a wafer, reaching 150k tokens/s for Llama 13B.
中科院“映天湖”原型集成16模组，性能改善1.45和1.78倍；Ouroboros架构单片集成54GB SRAM，Llama 13B推理达15万tokens/s。
New Ziguang Group's 'Zixian' 3D architecture delivered 30 TB/s memory bandwidth and is preparing for wafer-level integration.
新紫光集团“紫弦”3D架构实现30TB/s存储带宽，正着手晶圆级集成。