[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-ibm-z-arm-dual-isa-hot-chips-2026":3,"topics-all":35,"news-related-af79a195-07b7-4dae-8f7e-6e4a7e8b9d54":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"af79a195-07b7-4dae-8f7e-6e4a7e8b9d54","IBM 把 Arm 塞进大型机:Hot Chips 2026 拿出双指令集核心 + 第二代 AI 推理芯片","IBM 在 Hot Chips 2026 公布双指令集处理器:每颗核心原生执行 z\u002FArchitecture 与 AArch64,2nm 工艺、11 颗 5.7GHz 核心,搭配第二代 AI 推理加速器(16+1 冗余核心、FP4\u002FMXFP4、96GB HBM3e、PCIe Gen6)。","当一家做大型机做了六十多年的公司,愿意把 Arm 这种「非自家」的指令集以全硬件方式塞进核心——而且不是做两颗独立核心拼一块——这件事本身就值得停下来看一眼。IBM 在 Hot Chips 2026 上展示的新一代 IBM Z 与 LinuxONE 处理器,把这件事做实了。\n\n## 一颗核心跑两套 ISA,不是拼贴\n\n所谓「双指令集」(dual-ISA),指的就是每个核心既能原生执行 z\u002FArchitecture(IBM 大型机的老指令集),也能原生执行 AArch64——也就是大家熟悉的 Arm v9.3。注意关键词是「原生」:IBM 没有走翻译\u002F转译的捷径,而是把 AArch64 完整地实现在硬件上,小端 Arm 配大端 z\u002FArchitecture,并支持 SVE\u002FSVE2,实际实现的 AArch64 指令多达 2,792 条——按 IBM 自己在 Hot Chips 上的玩笑话,这数字比 z\u002FArchitecture 的指令还多,所以别再叫 RISC「精简」了。\n\n实现细节里有两个值得一提的取舍:一是直接消费 Arm 公开的 XML 架构描述来自动驱动解码器,分支预测的派发与发射逻辑则复用了 Z 核心的寄存器重命名,只是把 GR16-31 这段区间重新映射;二是把 AArch64 当作「一等公民」,而不是降级兼容层。结果是 Arm 软件无需修改、几乎可以「开箱即用」,而且 IBM 称其通过 Arm SystemReady 合规认证——这意味着可以直接跑标准 Arm 平台软件栈。\n\n这台处理器采用 2nm 工艺,单芯片 11 颗 IBM Z 高性能核心、主频 5.7GHz 以上、SMT=2,搭配每核 36MB 私有 L2、合并后形成 432MB 虚拟 L3 与 3.5GB 虚拟 L4 缓存,外加专用片上 DPU 做 I\u002FO 加速。可用性目标是 99.999999%(七个 9)。\n\n## 第二代 AI 推理芯片:堆叠「企业级」冗余\n\n很多人关注的是 AI 部分。这次 Hot Chips 上 IBM 还同步公布了第二代 AI 推理加速器,定位「企业级 GenAI」,参数列出来相当硬核:16 颗 AI 核心 + 1 颗冗余核心(可以容忍任意一颗核心失效),原生支持 FP4 与 MXFP4 数据类型,IBM 称 TOPS 最高可达上一代的 4 倍;96GB HBM3e 显存,峰值带宽约 4TB\u002Fs——大约是上一代的 20 倍;片间互联走 PCIe Gen6,以点对点低延迟方式与 IBM Z 主机对接。\n\n「冗余 AI 核心」这件事本身是个明显的信号:IBM 没把这颗加速器定位成训练卡或通用 GPU 的对手,而是把它当做一个面向银行交易、支付清算、保险理赔这种「不能停」的企业 AI 推理负载来设计的。和 z\u002FArchitecture 一样,它的安全栈也覆盖了机密计算与量子安全加密,做到数据与模型在静止、传输、使用三态都被保护。\n\n软件侧,IBM 把这些加速单元在 Linux on Arm 下暴露为标准 platform device,延迟与 z 原生指令基本相当;同时 z\u002FOS 还能继续把它们当原生指令用。这意味着同一台大型机上,IBM 老客户能继续跑 COBOL 与 s390x,而新写的 AI 应用可以直接基于 Arm 生态做,KVM\u002FOpenShift Virtualization 把两套工作负载塞进同一个逻辑分区——据称切换只在纳秒级。\n\n## 为什么这件事值得 AI 从业者关心\n\n把这件事和 OpenAI Jalapeño、英伟达 Vera 这类「通用推理芯片」摆在一起看会更有意思。OpenAI 走的是「针对自家模型体量优化 perf\u002FW」的路线,英伟达 Vera 押的是「数据中心级通用 CPU + GPU」组合拳;IBM 这条路则完全不同:把 AI 推理做成大型机的「原厂配件」,并且把「企业级可靠性 + 保密计算 + 量子安全」当成最关键的差异化卖点。\n\n这背后其实是在赌一件事:对于真正大额的金融、医疗、政府交易,大家最终愿意付钱的不是「跑得快多少倍」,而是「能不能不出事、能不能扛住监管审计、能不能和遗留系统在同一台机器上共存」。如果 IBM 这条路走通,OpenAI 与英伟达擅长的「互联网级」AI 推理,与 IBM 擅长的「关键任务级」AI 推理,会形成两套完全不同的工程范式。\n\n另一种解读更冷静:全球约 70% 的交易量仍跑在 IBM Z 上,这套双 ISA + AI 推理的组合,本质上是 IBM 给自家 60 年遗产续命的方案——把 Arm 软件生态引入大型机,把大型机的可靠性护城河延伸到 AI 推理,两边各取所需。至于「x86+GPU」「Arm 服务器+独立加速卡」之外的第三条路径能不能立住,要等首批客户实测数据出来再说。\n\n## 所以呢\n\n如果你正在做 AI 推理基础设施选型,IBM 这套方案短期内不会影响你的选型——它面向的不是云端大规模在线推理,而是高合规、强冗余、必须和遗留系统共存的私有部署场景。但它提供了一个清晰的信号:在大模型推理被「电力」「可靠性」「合规」三道墙限制的当下,把 AI 算力绑在「关键任务硬件」上,可能是一条和「堆更多 GPU」并行的工程路线。\n\n参考资料:\n- [ServeTheHome: IBM Z and LinuxONE Dual-ISA Processor and AI Acceleration at Hot Chips 2026](https:\u002F\u002Fwww.servethehome.com\u002Fibm-z-and-linuxone-dual-isa-processor-and-ai-acceleration-at-hot-chips-2026\u002F)\n- [Solidot: IBM 推出双指令集处理器](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85221)\n- [IBM Newsroom: IBM Z and LinuxONE Telum II \u002F Spyre](https:\u002F\u002Fchina.newsroom.ibm.com\u002F2026-08-26-IBM-IBM-Z-LinuxONE)","https:\u002F\u002Fwww.servethehome.com\u002Fibm-z-and-linuxone-dual-isa-processor-and-ai-acceleration-at-hot-chips-2026\u002F","745aa853-1fc9-4777-b7f9-f696c7d841fb",[11,15,18],{"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},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"73f3be5e-2397-4a46-b41c-38bd9777e455","en","IBM puts Arm inside the mainframe: Hot Chips 2026 ships a dual-ISA core plus a 2nd-gen AI inference chip","At Hot Chips 2026, IBM unveiled a dual-ISA processor: each core natively executes z\u002FArchitecture and AArch64 on a 2nm process, with 11 cores at 5.7GHz+. A second-gen AI inference accelerator pairs 16+1 redundant cores with FP4\u002FMXFP4 support, 96GB HBM3e, and PCIe Gen6.","When a company that has been building mainframes for more than sixty years is willing to put a non-native ISA like Arm into the core — and not as a sidecar — it deserves a closer look. At Hot Chips 2026, IBM showed the next-generation IBM Z and LinuxONE processor, and the design choices are anything but conventional.\n\n## One core, two ISAs — no stitching\n\nBy \"dual-ISA,\" IBM means every core can natively execute both z\u002FArchitecture (the mainframe's long-standing instruction set) and AArch64 (Arm v9.3). The keyword is \"native\": IBM did not take the translation or emulation shortcut. Instead, AArch64 is implemented directly in hardware, with a little-endian Arm sitting alongside a big-endian z\u002FArchitecture, plus SVE\u002FSVE2 support. In total, 2,792 AArch64 instructions are implemented — more, IBM quipped on stage, than the number of z\u002FArchitecture instructions, so stop calling RISC \"reduced.\"\n\nTwo implementation details stand out. First, the design ingests Arm's published XML architecture descriptions to feed the decoder automatically, while the dispatch and issue logic reuses the Z core's register-rename machinery, just remapped onto the GR16-31 range. Second, AArch64 is treated as a first-class citizen, not a compatibility afterthought. The result, IBM says, is that Arm software runs unmodified and out of the box, and that the core is Arm SystemReady-compliant — meaning it can absorb a standard Arm software platform as-is.\n\nThe chip is built on a 2nm process, packing 11 IBM Z high-performance cores running at 5.7GHz+, SMT=2, with a 36MB private L2 per core that aggregates into a 432MB virtual L3 and 3.5GB virtual L4 cache. A dedicated on-chip DPU handles I\u002FO acceleration. The availability target is 99.999999%.\n\n## Second-gen AI inference chip: redundancy as the differentiator\n\nMost of the AI attention, however, is on the accelerator IBM disclosed alongside the CPU. Positioned as \"enterprise GenAI,\" its spec sheet is unusually aggressive: 16 AI cores plus 1 redundant core (any single core failure is survivable), native FP4 and MXFP4 datatypes with TOPS up to 4x the previous generation, 96GB of HBM3e at roughly 4TB\u002Fs peak bandwidth (about 20x the prior generation), and PCIe Gen6 as a peer-to-peer low-latency host interface.\n\nThe \"redundant AI core\" choice is itself a tell. IBM is not positioning this part against training cards or general-purpose GPUs. It is built for the kind of AI inference that cannot fail: bank transactions, payment clearing, insurance adjudication. Like the rest of the Z stack, the accelerator is wrapped in confidential computing and quantum-safe cryptography, so data and models are protected at rest, in transit, and in use.\n\nOn the software side, IBM exposes these accelerators as standard platform devices under Linux on Arm, with latency comparable to native z-instructions; on z\u002FOS they remain first-class instructions. That means a single mainframe can keep running COBOL and s390x workloads for legacy customers while new AI applications sit on top of the Arm software stack. KVM and OpenShift Virtualization fold both worlds into the same logical partition, with thread-level switching reported at nanosecond scale.\n\n## Why AI practitioners should care\n\nSet this next to OpenAI's Jalapeño and NVIDIA's Vera, and the design space looks more interesting than usual. OpenAI is optimizing perf\u002FW for its own model volume; NVIDIA is building a CPU-plus-GPU combo for general-purpose data centers; IBM is betting on a different axis entirely — making AI inference a native peripheral of the mainframe, and selling \"enterprise-grade reliability plus confidential computing plus quantum-safe\" as the killer feature.\n\nUnderneath, this is a bet that for genuinely mission-critical financial, medical, and government transaction loads, customers will pay not for \"how many times faster\" but for \"does it stay up, can it survive an audit, and does it coexist with the legacy stack on the same box.\" If IBM's bet lands, the AI inference world will bifurcate: an \"internet-grade\" track tuned by OpenAI and NVIDIA, and a \"mission-critical-grade\" track where IBM sets the rules.\n\nA cooler reading is this: roughly 70% of global transaction volume still runs on IBM Z. The dual-ISA-plus-AI-accelerator combo is essentially IBM's plan to keep that sixty-year-old franchise alive — bringing the Arm software ecosystem into the mainframe and extending the mainframe reliability moat into AI inference, with each side taking what the other needs. Whether the third path — beyond \"x86+GPU\" and \"Arm server+discrete accelerator\" — actually holds, will depend on what the first wave of customers measure in production.\n\n## So what\n\nIf you are choosing an AI inference stack today, IBM's new offering will not affect your decision in the short term — it targets private deployments with strict compliance, redundancy, and legacy coexistence requirements rather than large-scale cloud inference. But it sends a clear signal: as large-model inference runs into the walls of power, reliability, and compliance, bolting AI compute onto \"mission-critical hardware\" may run in parallel with the \"stack more GPUs\" playbook rather than competing with it head-on.\n\nReferences:\n- [ServeTheHome: IBM Z and LinuxONE Dual-ISA Processor and AI Acceleration at Hot Chips 2026](https:\u002F\u002Fwww.servethehome.com\u002Fibm-z-and-linuxone-dual-isa-processor-and-ai-acceleration-at-hot-chips-2026\u002F)\n- [Solidot: IBM unveils dual-ISA processor](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85221)\n- [IBM Newsroom: IBM Z and LinuxONE](https:\u002F\u002Fchina.newsroom.ibm.com\u002F2026-08-26-IBM-IBM-Z-LinuxONE)","ibm-z-arm-dual-isa-hot-chips-2026","2026-08-28T11:00:00Z","2026-08-28T11:03:21.408881Z","2026-08-28T11:03:21.408893Z",true,"agent",156,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"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":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"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":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"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":62,"title":63,"news_slug":64,"published_at":65},"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":67,"title":68,"news_slug":69,"published_at":70},"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":72,"title":73,"news_slug":74,"published_at":75},"1bf3f192-68e9-49f1-8bbb-a350f05c15e2","LLM 编排云边端:FRAME 2026 论文点出反馈层缺环","llm-cloud-continuum-frame-2609","2026-09-16T11:00:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"367476a4-b9af-46f1-a6ab-3de1d83640ff","NVIDIA 把中国开发者日搬到苏州:10 月连开两天,AI 推理和物理 AI 是主菜","nvidia-china-developer-day-2026-suzhou","2026-09-16T03:00:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"b398dc77-58a1-498e-a9ed-c045c83c90be","AI 抢走消费级 DRAM:一年涨价五倍,手机路由器全被拖下水","ai-dram-consumer-electronics-price-surge","2026-09-14T01:00:00+00:00"]