[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-ibm-z-arm-dual-isa-hot-chips-aug-2026":3,"topics-all":38,"news-related-5d452086-ecb7-494d-82b9-d962664aa243":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"5d452086-ecb7-494d-82b9-d962664aa243","IBM 把 Arm 核塞进 Z 大型机:Hot Chips 2026 公布业界首款双指令集处理器","IBM 在 Hot Chips 2026 发布业界首款双指令集大型机处理器,采用 2nm 工艺、11 颗主频超过 5.7 GHz 的核心,内置面向交易欺诈的 AI 推理加速器。每个核心可原生执行 Arm 与 IBM Z 指令,把 Arm 生态带入关键任务型企业基础设施。","全球约七成信用卡和金融交易目前仍在 IBM Z 大型机上跑完。这台被打上「关键任务型企业基础设施」标签的硬件,8 月 26 日在 Hot Chips 2026 大会上拿到了二十多年来最大的一次架构重写:IBM 与 Arm 联合公布了业界首款可以在同一颗核心上原生执行两套指令集的处理器,过去那种通过模拟或独立小核桥接 Arm 生态的方案被直接淘汰。\n\n## 一颗核心、两套指令:不是「1+1」,是「直接合并」\n\n处理器采用 2 纳米工艺节点,封装 11 颗主频超过 5.7 GHz 的高性能核心。每一颗核心都可以同时执行 Arm 和 IBM Z(以及 Arm 和 LinuxONE)指令集,无需借助 Arm 与 IBM 核心并存的双核方案。IBM 院士、系统开发业务首席技术官 Christian Jacobi 在发布稿中表示,这种「在架构层面」的合并让企业可以在 IBM Z 上同时跑 z\u002FOS、Linux on IBM Z 以及 Arm 原生 Linux 环境,「既帮助客户接入业界增长最快的软件生态之一,也延续了 IBM 系统作为当今企业运营基石的核心品质」。([ibm.com](https:\u002F\u002Fchina.newsroom.ibm.com\u002F2026-08-26-IBM-IBM-Z-LinuxONE))\n\n传统上,把 x86、Arm、RISC-V 等指令集塞进一台机器,惯用做法是「各自一颗核」:Intel 的 Lakefield 与 AMD 的 Kria 板卡都通过小核桥接配系统唤醒来做能效与兼容性平衡。IBM 这次直接放弃物理分核,要求每颗核心在解码阶段就用硬件同时认两套指令字。技术报告里讲得很直接:这种做法对前端解码器和分支预测器提出了更高要求,但换来的是单线程可以灵活切换工作负载、对调用者完全透明。Solidot 援引 Hot Chips 议程指出,IBM Z 平台可扩展到「数百个核心和数十 TB 内存」,这种「全部核心都能跑 Arm」的拓扑,意味着 Linux 容器化与 z\u002FOS 事务负载第一次可以放在完全一致的硬件资源池里抢占调度。([solidot.org](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85221))\n\n## 为什么塞 5.7 GHz 的 Arm 进主机?AI 推理加速器是答案\n\nIBM 没有把这件事讲成「我支持 Arm 了」那么简单。发布材料里同步强调了一项直接面向业务的硬件单元:**面向交易过程中欺诈检测的 AI 推理加速器**。在金融行业,反欺诈模型正在从离线批处理走向实时打分;每笔交易在毫秒级窗口内,需要把行为序列、向量数据库召回结果与风控模型合并推理。把这类负载交给主机 CPU 跑,既挤占原本用来处理核心账务的周期,又撑不起延迟 SLA;走外部 GPU 加速器,则要跨 PCIe、跨 NUMA、再跨可信边界。\n\nIBM 的解法是把推理加速器做到片上,与 11 颗 5.7 GHz 核心共享大容量缓存架构,I\u002FO 环节也由专用片上数据处理单元分担。这意味着欺诈评分可以直接吃主机的内存语义,既不需要 ETL 同步、也不需要把交易数据搬到外部推理集群。Arm 云 AI 事业部执行副总裁 Mohamed Awad 在发布稿里直接点出了背后逻辑:「随着 AI 的规模化应用,越来越多的计算场景正聚集于 Arm 架构……将 Arm 的计算能力和软件生态引入 IBM 平台,将进一步推动这一趋势深入关键任务型企业基础设施,为企业的 AI 部署方式提供更多选择。」([ibm.com](https:\u002F\u002Fchina.newsroom.ibm.com\u002F2026-08-26-IBM-IBM-Z-LinuxONE))\n\n## 这是 Arm 的企业级背书,也是主机厂商的「绑死」策略\n\n值得关注的另一条时间线:2026 年 4 月,IBM 与 Arm 正式公布战略合作,8 月这颗芯片就是合作条款里写明要在年内交付的「首个里程碑式成果」。在 Arm 一侧,这等于拿下了一份跨金融、政府、保险等「高度监管行业」的标杆案例——这恰恰是 x86 生态最难啃的板块。在 IBM 一侧,这是把主机业务从「IBM Z 专属语言 + 模拟跑 Linux」这条曲线,直接拉到「Arm 原生云原生 + z\u002FOS 兼容」的双轨通道。([china.newsroom.ibm.com](https:\u002F\u002Fchina.newsroom.ibm.com\u002F2026-08-26-IBM-IBM-Z-LinuxONE))\n\n对于正在做核心现代化的大型机构(银行、券商、医保清算、税务系统等),这是一条值得关注的迁移信号:他们原本必须为「跑 Java\u002FPython 容器栈」专门搭一套 x86 或 LinuxONE 后端;以后这部分负载可以直接放在 IBM Z 上,既享受主机的可靠性、加密、故障恢复能力,又把 Arm 上现成的 AI\u002FML 工具链直接拉过来,容器镜像、Kubernetes operator、PyTorch 都不需要重写。对于计划在主机侧做推理业务的 ISV,这意味着过去只能跑 SVM 或规则引擎的交易风控栈,现在有机会切换成端到端的大模型评分,而延迟和合规边界由 IBM Z 兜底。\n\n## 写在最后\n\n「双指令集 + 5.7 GHz + AI 推理加速器」三个标签放在一起,本质上是 IBM 给 Z 系列装上了一个「不会让 Arm 生态长大了之后再回头掐自己脖子」的接口。当 AWS Graviton、NVIDIA Grace 等 Arm 服务器用性价比把大量云原生工作负载拉走的时候,IBM 选择了不争云端市场,而是把 Arm 当成扩展主机生态的杠杆。下一阶段真正值得盯的指标,不是 IBM Z 的出货量,而是 Arm 基金会拿到这份「关键任务级 + 实时 AI」订单后,会在多大程度上影响微软 Azure Cobalt、英伟达 Grace 等其他 Arm 服务器的路线图。","https:\u002F\u002Fchina.newsroom.ibm.com\u002F2026-08-26-IBM-IBM-Z-LinuxONE","653dda08-2edc-4d17-aeb2-56b0c88dd918",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"fca9258a-9430-455a-b95d-b9fae5e373a8","ai-inference",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"74948093-90af-4359-86a1-8a7911cee397","en","IBM Squeezes Arm Cores Into Z Mainframes: Industry-First Dual-ISA Processor Revealed at Hot Chips 2026","At Hot Chips 2026, IBM unveiled the industry's first dual-ISA mainframe processor. Built on a 2nm process with 11 cores clocked above 5.7 GHz, it bakes in an AI inference accelerator aimed at transaction-time fraud detection. Every core natively executes both Arm and IBM Z instructions, pulling the Arm ecosystem into mission-critical enterprise infrastructure.","About seventy percent of the world's credit card and financial transactions still complete on IBM Z mainframes. On August 26, at Hot Chips 2026, that flagship of \"mission-critical enterprise infrastructure\" got its biggest architectural rewrite in over two decades: IBM and Arm jointly unveiled the industry's first processor capable of natively executing two instruction sets on the same core, retiring the era of emulation bridges or side-car Arm cores.\n\n## One core, two ISAs: not \"1+1,\" but \"directly fused\"\n\nThe chip is built on a 2nm process node and packages 11 high-performance cores running above 5.7 GHz. Each core can execute both the Arm and IBM Z (as well as Arm and LinuxONE) instruction sets natively, with no need for a dual-core arrangement that keeps separate Arm and IBM cores side by side. In the announcement, IBM Fellow and CTO of Systems Development Christian Jacobi framed the move: this \"architectural-level\" merge lets enterprises run z\u002FOS, Linux on IBM Z, and Arm-native Linux workloads on IBM Z at the same time, \"giving customers access to one of the fastest-growing software ecosystems in the industry while preserving the core qualities that make IBM systems the foundation of today's enterprise operations.\" ([ibm.com](https:\u002F\u002Fchina.newsroom.ibm.com\u002F2026-08-26-IBM-IBM-Z-LinuxONE))\n\nHistorically, squeezing x86, Arm, or RISC-V into the same machine meant \"one core per ISA\": Intel's Lakefield and AMD's Kria boards balanced efficiency and compatibility through small cores bridged via system calls. IBM threw out the physical split and demanded that every core recognize both instruction streams in its front-end decoder. The technical report is blunt: the approach raises the bar for the front-end decoder and branch predictor, but pays back in single-thread flexibility and full transparency to the caller. Citing the Hot Chips agenda, Solidot adds that IBM Z platforms scale to \"hundreds of cores and tens of terabytes of memory.\" With \"every core running Arm,\" Linux container and z\u002FOS transaction workloads can finally be scheduled inside the exact same hardware resource pool. ([solidot.org](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85221))\n\n## Why push 5.7 GHz Arm into a mainframe? The AI inference accelerator is the answer\n\nIBM did not pitch this as a simple \"we support Arm now.\" The release materials spotlight one business-facing hardware block: **an AI inference accelerator aimed at fraud detection during transactions**. In financial services, anti-fraud models are migrating from offline batch scoring to real-time scoring. Each transaction, inside a millisecond window, must combine behavioral sequences, vector-database recall output, and the risk model. Routing that load to the host CPU eats cycles that were reserved for core bookkeeping and still cannot meet the latency SLA. External GPU accelerators, on the other hand, mean crossing PCIe, crossing NUMA, and crossing a trust boundary.\n\nIBM's remedy is to push the inference accelerator on-die, sharing a high-capacity cache fabric with the 11 cores clocked above 5.7 GHz, while a dedicated on-chip data-processing unit takes over I\u002FO. Fraud scoring can therefore consume the mainframe's memory semantics directly, with no ETL staging, no transaction-data shuttle to an external inference cluster. Arm Executive Vice President of Cloud AI Mohamed Awad spelled out the logic in the release: \"As AI scales out, more and more compute is congregating on the Arm architecture … bringing Arm's compute and software ecosystem onto IBM platforms will push this trend deeper into mission-critical enterprise infrastructure and give enterprises more choices for how they deploy AI.\" ([ibm.com](https:\u002F\u002Fchina.newsroom.ibm.com\u002F2026-08-26-IBM-IBM-Z-LinuxONE))\n\n## Enterprise validation for Arm, a lock-in play for the mainframe vendor\n\nAnother timeline worth tracking: in April 2026, IBM and Arm formally announced their strategic partnership, and this chip is the \"first milestone\" their agreement committed to deliver within the year. For Arm, the deal is a marquee reference across banking, government, insurance, and other \"heavily regulated verticals\" — precisely the segment that is hardest for the x86 ecosystem to crack. For IBM, it shifts the mainframe business off the \"IBM-Z-only language plus emulated Linux\" curve and straight onto a dual track of \"Arm-native cloud-native\" plus \"z\u002FOS-compatible.\" ([china.newsroom.ibm.com](https:\u002F\u002Fchina.newsroom.ibm.com\u002F2026-08-26-IBM-IBM-Z-LinuxONE))\n\nFor large institutions running core modernization programs (banks, brokerages, medical-insurance clearinghouses, tax systems), this is a migration signal worth watching: today they have to spin up a separate x86 or LinuxONE backend just to host their Java\u002FPython container stack. Going forward, that workload can sit on IBM Z, inheriting the mainframe's reliability, encryption, and fault-recovery guarantees while pulling in Arm's full AI\u002FML toolchain. Container images, Kubernetes operators, PyTorch — none have to be rewritten. For ISVs planning to do inference on the mainframe, this means the SVM- and rule-engine-only fraud stacks of the past can now be re-implemented as end-to-end large-model scoring, with IBM Z holding the latency and compliance line.\n\n## Closing thoughts\n\nStack \"dual-ISA + 5.7 GHz + AI inference accelerator\" together, and the takeaway is that IBM has bolted onto the Z line an interface that lets the Arm ecosystem grow up without, years later, turning around and squeezing the mainframe vendor's throat. As AWS Graviton and NVIDIA Grace use price-performance to drag cloud-native workloads onto Arm, IBM has chosen not to fight for the cloud market but to use Arm as a lever for extending the mainframe ecosystem. The numbers worth watching next are not IBM Z shipments but the route maps those numbers pull on at Microsoft (Azure Cobalt), NVIDIA (Grace), and the rest of the Arm-server field once Arm Holdings lands a \"mission-critical + real-time AI\" reference of this size.","ibm-z-arm-dual-isa-hot-chips-aug-2026","2026-08-29T06:00:00Z","2026-08-29T11:03:17.086046Z","2026-08-29T11:03:17.086053Z",true,"agent",267,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"8f244da3-58a0-430f-9c57-af8df68a1337","高通把数据中心 HBC 架构塞进手机：2028 年商用,端侧 LLM 推理的「内存墙」破局战","qualcomm-hbc-phone-2028-ondevice","2026-06-29T02:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"851e2c6d-4de9-4477-a80b-b06cf16b12b6","iOS 27 端侧 Apple Intelligence 系统级整合：1H27 新机 DRAM 升级至 9GB，是硬件先行的信号","ios27-apple-intelligence-dram-9gb","2026-06-28T02:28:46+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"cb1e799d-d6d7-4ab9-9eaf-bea0aa432b06","Mistral 模型进驻 Firefox:119B 开放权重模型驱动浏览器 AI 助手","mistral-small-4-firefox-smart-window","2026-09-16T17:07:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"b48106fa-b9c7-4317-a40b-61e4219a698a","IBM Z 双架构处理器:把 Arm 推进大型机,内置 AI 推理","ibm-z-dual-architecture-arm-ai-inference","2026-09-03T03:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"97eb895f-8294-47b9-a9e0-5f165a75cb5b","OpenAI Jalapeño 实测:自研推理芯片在 Hot Chips 上跑赢 Blackwell","openai-jalapeno-hot-chips-broadcom-blackwell","2026-08-26T08:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"4aa9534a-778e-4cd7-8194-fdf3097249b8","OpenAI Jalapeño Hot Chips 实测:峰值每瓦 1.9×,延迟压到 1 秒","openai-jalapeno-hot-chips-benchmark-2026","2026-08-26T02:00:00+00:00"]