当前位置:首页 > 报告详情

11.30 Hussam Amrouch - TUM.pdf

上传人: sec****ies 编号:734586 2025-07-26 28页 3.41MB

1、1by H.Amrouch Thermal Management Expo in Stuttgart E-mail:amrouchtum.deChair of AI Processor Design,Technical University of MunichConfidentialChallenges in the Era of AI Chips and Advance Technologies by Hussam AmrouchChair of AI Processor Design2by H.Amrouch Thermal Management Expo in Stuttgart E-m

2、ail:amrouchtum.deChair of AI Processor Design,Technical University of MunichConfidentialAI:The Next RevolutionComputing demand3.4 months doubling!Source:https:/1e+4Petaflop/s-day1e+21e01e-21e-41e-61e-81e-101e-121e-141960197019802000201020201990Modern Era First Era3by H.Amrouch Thermal Management Exp

3、o in Stuttgart E-mail:amrouchtum.deChair of AI Processor Design,Technical University of MunichConfidentialDeep Learning Breakthroughs in the AccuracyIts possible because Efficient AI ChipsAI Chip:Google TPUv1 ISCA17Complex DNN on one TPUv3:1.8min 2048 GPUs+512 CPUspictures sources:by GDJ,openclipart

4、.org and https:/ H.Amrouch Thermal Management Expo in Stuttgart E-mail:amrouchtum.deChair of AI Processor Design,Technical University of MunichConfidentialIts possible because Efficient AI ChipsBUT.More Efficiency is really Good?5by H.Amrouch Thermal Management Expo in Stuttgart E-mail:amrouchtum.de

5、Chair of AI Processor Design,Technical University of MunichConfidentialLesson from Previous Revolution?src:https:/learnodo- go back to 1865src:WikipediaWilliam JevonsJevons ParadoxWhen technology increases theefficiency,the consumption rises Gain from efficiency will backfire!6Confidentialby H.Amrou

6、ch Thermal Management Expo in Stuttgart E-mail:amrouchtum.deChair of AI Processor Design,Technical University of MunichAI Acceleration and Efficiency ParadoxMore Efficient HW for AI Larger and larger AI models Memory Bottleneck!Significant Efficiency Loss Novel architectures are MUST7by H.Amrouch Th

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。
本文主要讨论了AI芯片时代面临的挑战及其技术进步。关键点如下: 1. AI计算需求激增,每3.4个月翻一番。 2. 高效AI芯片如谷歌TPU使得深度学习突破成为可能,但高效硬件可能导致模型越大,内存瓶颈和效率损失越显著。 3. AI硬件面临能耗、散热和内存限制问题,如电压、内存和冷却墙。 4. 产业需要创新,例如在7nm技术节点中,自热和可靠性是主要问题。 5. 提出了新型冷却技术,如超级晶格热电冷却,实现局部、按需高效冷却。 6. 智能温度感应和冷却技术可实时获取热图,为处理器芯片提供按需冷却。 核心数据引用:谷歌TPU的运算能力、深度学习模型的功耗需求、7nm纳米片的自热数据、热电冷却效率,以及热传感器阵列的密度等。
"效率提升会带来什么后果?" "AI芯片的散热难题如何解?" "未来AI芯片的创新方向是?"
客服
商务合作
小程序
服务号
折叠