当前位置:首页 > 报告详情

SemiAnalysis-GTC 2026 – 推理王国扩张 GTC 2026 – The Inference Kingdom Expands(69页).pdf

上传人: 山海 编号:1170946 2026-03-31 69页 38.71MB

下载:

1、Nvidia The Inference Kingdom Expands英伟达 推理王国扩张Groq LP30,LPX Rack,Attention FFN Disaggregation,Oberon&Kyber Updates,Nvidias CPO Roadmap,Vera ETL256,CMX&STXGroq LP30、LPX 机架、Attention FFN 解耦、Oberon 与 Kyber 更新、英伟达 CPO 路线图、Vera ETL256、CMX 与 STXDYLAN PATEL,MYRON XIE,DANIEL NISHBALL,AND 7 OTHERSDYLAN PATEL

2、,MYRON XIE,DANIEL NISHBALL 以及其他 7 位作者MAR 24,2026 2026 年 3 月 24 日 PAID 付费内容1Share 分享Source:Nvidia 来源:英伟达At GTC 2026,Nvidia delivered an event packed full of ground breakingannouncements.Nvidias pace of innovation is not showing any signs of slowing,as27they introduced three entirely new systems this

3、year:Groq LPX,Vera ETL256,andSTX.Also announced were updates to Nvidias Kyber rack architecture system,CPOmaking its debut for scale-up networking with the unveiling of the Rubin UltraNVL576 and Feynman NVL1152 multi-rack systems.Early hints on Feynmansarchitecture was also a key topic.A Jensen call

4、out for InferenceX during the keynotewas a highlight.在 GTC 2026 上,英伟达(Nvidia)带来了场充满突破性发布的盛会。英伟达的创新步伐没有表现出任何放缓的迹象,今年他们推出了三款全新的系统:Groq LPX、Vera ETL256 和 STX。同时发布的还有英伟达 Kyber 机架架构系统的更新,随着Rubin Ultra NVL576 和 Feynman NVL1152 多机架系统的亮相,CPO(共封装光学)在扩展络领域次登场。关于 Feynman 架构的早期线索也是个核话题。仁勋在主题演讲中对 InferenceX 的点名表

5、扬成为了亮点。This is our GTC 2026 recap,and we will address many of the key questions that havebeen left unanswered by Nvidia.Specifically,we will go through the LPX rack andLP30 chip and explain how attention and feed forward network disaggregation(AFD)works;more details on the various rack architectures

6、 behind NVL144,NVL576,andNVL1152 and clarify just how much optics will be inserted as well as the rationalebehind the dense Vera ETL256.The next generation Kyber rack had some big updatesand some hidden details.这是我们的 GTC 2026 回顾,我们将解答英伟达留下的许多关键问题。具体,我们将深探讨 LPX 机架和 LP30 芯,并解释注意机制与前馈络解耦(AFD)的作原理;详细介绍

word格式文档无特别注明外均可编辑修改,预览文件经过压缩,下载原文更清晰!
三个皮匠报告文库所有资源均是客户上传分享,仅供网友学习交流,未经上传用户书面授权,请勿作商用。
1. **英伟达GTC 2026发布三大新系统**:Groq LPX机架、Vera ETL256(256 CPU/机架,液冷)、STX存储机架,并更新Kyber架构(Rubin Ultra NVL144/NVL576)及CPO路线图(Rubin Ultra NVL576机架间用CPO)。 2. **Groq LPU集成与AFD技术**:LP30芯片(500MB SRAM,1.2 PFLOPs FP8)与GPU协同,通过注意力与前馈网络解耦(AFD)优化推理延迟,LPX机架含32个1U计算托盘(16 LPU/托盘)。 3. **网络与架构升级**:Kyber机架密度提升(144 GPU/机架),NVL576采用8个Oberon机架+CPO互连;Feynman规划NVL1152(8 Kyber机架),可能全CPO。 4. **存储与计算扩展**:CMX平台(KV缓存卸载)+STX参考架构(32 Vera CPU/机架),BlueField-4 DPU推动存储标准化。 5. **供应链影响**:Qualcomm(AlphaWave)提供LP30 SerDes;LP40将转台积电N3P工艺,集成NVLink。
**LPU如何提升推理性能?** **CPO技术如何改变网络架构?** **Vera ETL256为何采用液冷?**
客服
商务合作
小程序
服务号
折叠