5AI Subscription vs. H100 [video](youtube.com)📰 HN💡 The title focuses on comparing an AI subscription with Nvidia H100 hardware, indicating a discussion or debate topic rather than a product launch, research paper, or business update. 7Furiosa: 3.5x efficiency over H100s(furiosa.ai)📰 HNFuriosaAI推出了NXT RNGD服务器，这是一款基于其RNGD加速器的一站式AI推理解决方案。它为关键AI工作负载提供高性能，可无缝集成到现有数据中心，预装了Furiosa SDK和LLM运行时以便应用安装后立即服务，并采用标准PCIe互连以避免专有基础设施。 9Data on AI Chip Sales(epoch.ai)📰 HNEpoch AI发布了一个AI芯片销售公开数据库，估算了英伟达、AMD、谷歌、华为等主要芯片设计商销售的专用AI加速器数量。该数据集包含芯片型号细分信息以及计算能力（以H100等效值衡量）等指标，不同芯片设计商的估算置信度存在差异。 14Who Invented the Transistor?(people.idsia.ch)📰 HN本文探讨了晶体管发明的历史，强调Julius Edgar Lilienfeld在1925-28年申请的场效应晶体管（FET）专利是现代计算机和智能手机中使用的基础设计。文章将其与贝尔实验室1948年的点接触晶体管（后者为死路）进行对比，并提及后续的FET变体以及Lilienfeld与贝尔实验室之间的优先权争议。 16After the Bubble(tbray.org)📰 HN本文分析了生成式AI泡沫即将破裂的问题，重点指出GPU的脆弱性（如Llama 3训练期间Nvidia H100的故障）和高功耗成本是关键因素。文章提到，与过去的泡沫（铁路、互联网泡沫）破裂后留下有价值基础设施不同，由于GPU损耗快和能源成本高，生成式AI泡沫破裂后可能不会留下类似的长期价值，并指出特殊目的实体（SPVs）是大型科技公司在不增加资产负债表债务的情况下建设AI数据中心的财务手段。 23Deploying DeepSeek on 96 H100 GPUs(lmsys.org)📰 HNLMSYS团队使用SGLang在96块H100 GPU（12个节点×8）上部署了DeepSeek大语言模型，采用预填充-解码分离和大规模专家并行技术。该实现达到了高吞吐量（对于2000 token输入，每个节点每秒处理52.3k输入token和22.3k输出token），性能与DeepSeek官方报告相当，成本仅为其API的五分之一，且完全开源并提供可复现的实验指导。 24The Future of Compute: Nvidia's Crown Is Slipping(mohitdagarwal.substack.com)📰 HN💡 The story discusses Nvidia's slipping dominance in the compute market, which is directly linked to AI hardware (e.g., GPUs like H100 used for AI training/inference), aligning with the hardware category. 27CUDA Moat Still Alive(semianalysis.com)📰 HN💡 The story discusses benchmarks of AI chips (Nvidia H100/H200, AMD MI300x) and CUDA's competitive advantage, which are core to AI hardware and compute. 29U.S. chip revival plan chooses sites(spectrum.ieee.org)📰 HN💡 The story about the U.S. chip revival plan choosing sites relates to semiconductor manufacturing infrastructure, which is critical for AI compute (e.g., chips like Nvidia H100 used in AI data centers). 30Ultraprecise method of aligning 3D semiconductor chips invented(techxplore.com)📰 HN💡 The story focuses on an ultraprecise method for aligning 3D semiconductor chips, which are key components in AI hardware like Nvidia's H100 (using 3D stacking). This directly relates to AI-related chips and compute, fitting the hardware category. 📅周报