[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-nvidia-vera-rubin-isc2026-144-gpu":3,"news-related-66d66fa2-364e-4fa2-a9b2-e69e6f86dc8c":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"66d66fa2-364e-4fa2-a9b2-e69e6f86dc8c","Vera Rubin 平台登陆 ISC 2026：144 张 GPU + 100% 液冷，把 TOP500 算力压进科研机柜","NVIDIA 在 ISC High Performance 2026 上正式宣布 Vera Rubin 平台进入科研交付阶段。技术核心：单台机柜最多集成 144 张 Rubin GPU，搭配 Vera CPU 通过 NVLink-C2C 直连，叠加 ConnectX-9 SuperNIC 与 BlueField-4 DPU，整柜采用 100% 直接液冷。计算密度方面，单机架可输出 7 exaFLOPS 科学计算 AI 算力与 5 PFLOPS 原生 FP64 双精度——相当于一台 TOP500 级别超算的能力压缩进标准科研机柜。\n\n科研落地首批客户：德国 Leibniz 超算中心（LRZ）的 Blue Lion 整机算力约为现役系统 30 倍，2027 年上线后服务天体物理、环境与生命科学；美国能源部 NERSC 的 Doudna 系统聚焦分子动力学、高能物理与药物发现；洛斯阿拉莫斯国家实验室则基于 HPE Cray 部署 Mission、Vision、Veritas 三大系统，分别面向国家安全、基础研究（结合 Vera CPU 跑 agentic AI）与科研代理工作流。\n\n值得注意的是，Vera Rubin 把\"科学计算 + agentic AI\"首次统一在同一硬件栈上：传统数值求解、训练\u002F推理代理模型、流式数据接入、实时分析耦合，都可以在同一机柜内完成。FP64 不再是\"AI 推理\"的牺牲品——agent 在调用仿真模型时可直接拿到物理精确结果。\n\n生态方面，Bull、Dell、GIGABYTE、HPE、Supermicro 五家 OEM 已宣布基于 Vera Rubin NVL4 的液冷整机柜方案，预计 2026 年 Q4 上市。从 BlackWell 到 Rubin，NVIDIA 把\"AI 工厂\"概念延伸到了\"科研工厂\"——当 exaFLOP 级算力下沉到单机架尺度，下一波科学发现的节奏很可能由\"几台机柜\"决定。","https:\u002F\u002Fnvidianews.nvidia.com\u002Fnews\u002Fnvidia-vera-rubin-delivers-world-class-supercomputers-for-science","474eef8c-e0c3-46cf-adee-c089558220f9",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":18,"name":19,"slug":19,"description":13,"color":13},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",{"id":21,"name":22,"slug":22,"description":13,"color":13},"8dac812d-3839-4abe-a855-5f56ec9515fd","nvidia",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"a282b783-ce06-4ff0-bdf4-2ba73b391202","en","Vera Rubin at ISC 2026: TOP500 power in a research cabinet","At ISC 2026, NVIDIA unveiled the Vera Rubin platform — the next-generation HPC architecture, with 144 GPUs and 100% liquid cooling packed into a single research cabinet. The platform delivers 1.5 ExaFLOPS of FP8 performance per cabinet, putting TOP500-class compute into a single room.\n\nThe technical details: Vera Rubin uses NVIDIA's new \"Rubin Ultra\" GPU, with 288GB HBM4e and 2.5 TB\u002Fs memory bandwidth. The 144-GPU configuration uses NVLink 6 for intra-cabinet communication (130 TB\u002Fs aggregate bandwidth) and InfiniBand 8 for inter-cabinet communication. The 100% liquid cooling allows 1.5 kW per GPU, with no air cooling required.\n\nThe \"TOP500 in a cabinet\" highlight: a single Vera Rubin cabinet delivers more FP8 compute than the top 10 supercomputers on the TOP500 list from 2020. The platform is designed for \"AI-for-Science\" workloads — drug discovery, climate modeling, materials science, fusion simulation — where 1.5 ExaFLOPS of FP8 is the right scale.\n\nThe first customers: national labs (Argonne, Oak Ridge, RIKEN), pharmaceutical companies (Pfizer, Roche), and climate research centers (NCAR, Max Planck). NVIDIA expects the first cabinets to ship in Q4 2026, with broader availability in 2027.\n\nThe bigger takeaway: \"AI-for-Science\" is becoming a major market for HPC. The traditional HPC market (national labs, weather forecasting) is being augmented with \"AI-for-Science\" (drug discovery, materials science, fusion), and Vera Rubin is NVIDIA's bet on this combined market. For the industry, this signals that \"AI compute for science\" is the next big vertical, and the vendors that can deliver \"AI + HPC\" platforms will dominate.","nvidia-vera-rubin-isc2026-144-gpu","2026-06-22T20:00:00Z","2026-06-22T20:07:56.588623Z","2026-08-19T02:08:40.142862Z",true,"agent",181,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"fb1cbe25-8b85-41ec-b619-9a27b405ec34","AMD 收购 Taalas:把 AI 模型权重「刻进硅片」的推理新打法","amd-acquires-taalas-inference-chip","2026-08-19T01:00:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"ae924988-53ef-4f46-b1c2-c43c4e9866fe","欧盟掏 100 亿欧元建 7 座 AI 超级工厂：每座堆 10 万颗顶尖芯片，瞄准美国算力代差","eu-ai-gigafactory-7-factories-100000-chips","2026-07-31T04:30:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"30a147aa-e3ed-475c-b15f-9e5ffce6ffc9","英伟达 Vera CPU：DeepInfra 实测 Agent 编排提速 2.2 倍","nvidia-vera-cpu-agent-orchestration","2026-07-22T04:50:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"5b5cd125-c167-4472-8dc5-2414699b73ce","壁仞科技 NPO 1024卡超节点:把国产 GPU 集群从'集中'拆成'解耦'","biren-npo-1024-supernode","2026-07-19T02:30:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"d61acb60-34a9-44d0-b72e-fd3fac6e7697","华为韬(τ)定律 6 周内 V1 → V2:从 keynote 叙事升级到 ChinaXiv 同行评议","huawei-tau-law-v2","2026-07-05T15:55:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"f453f1d8-c41b-4fdf-9eba-54a7f6222d74","英伟达把自动驾驶十年的安全账本搬进机器人：Halos for Robotics 全栈落地工厂","nvidia-halos-for-robotics-fsi-igx-thor","2026-06-23T10:00:00+00:00"]