[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-openai-jalapeno-hot-chips-benchmark-2026":3,"news-related-4aa9534a-778e-4cd7-8194-fdf3097249b8":41},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"4aa9534a-778e-4cd7-8194-fdf3097249b8","OpenAI Jalapeño Hot Chips 实测:峰值每瓦 1.9×,延迟压到 1 秒","Hot Chips 2026 上 OpenAI 披露 Jalapeño 推理芯片实测:在 SemiAnalysis InferenceX 上对比 Nvidia Blackwell,峰值吞吐\u002F瓦 1.5–1.9×,端到端延迟低 1.7–3.6×,每用户 tokens\u002Fs 最高 1,459(GB200 仅 535)。","OpenAI 的「自研推理芯片」传闻从去年讲到现在,但 8 月 25 日 Hot Chip 2026 上,Richard Ho 团队第一次把数字摆到台面上。这是 OpenAI 第一次认真回答两个问题:第一,自研硅究竟比 Nvidia Blackwell 快多少;第二,九个月就跑通 tape-out 的工程奇迹,会不会卡在「AI 自己写 kernel」这一步上。\n\n## 三张图表里的真实增量\n\nOpenAI 没有选自己最擅长的内部模型,而是拿三个公开模型在 SemiAnalysis 的 InferenceX 基准上跑测试,这样任何人都能复现:\n\n- **GPT-OSS 120B**:峰值吞吐 85,448 vs 44,960 mixed tokens\u002Fs per kW(对比 GB200),即 **1.9 倍**;端到端延迟从 1.80 秒降到 1.03 秒。\n- **DeepSeek R1 670B(MXFP4)**:19,641 vs 11,781 mixed\u002FkW,**1.7 倍**;延迟 1.65s vs 5.99s。\n- **Kimi K2.5 1T(MXFP4)**:18,195 vs 11,862 mixed\u002FkW,**1.5 倍**;延迟 1.56s vs 5.31s。\n\n更直观的是单用户感受。在 DeepSeek R1 上,Jalapeño 每秒给单个用户吐 **700 tokens**,对比系统只有 169;Kimi K2.5 上是 **694 vs 182**;GPT-OSS 上 **1,459 vs 535**。换句话说:用户感知到的「卡不卡」,差距最高拉到 **4 倍以上**。\n\n来源:[OpenAI 官方披露页](https:\u002F\u002Fopenai.com\u002Findex\u002Fjalapeno-first-results\u002F)与 [TechCrunch 报道](https:\u002F\u002Ftechcrunch.com\u002F2026\u002F08\u002F25\u002Fopenais-jalapeno-chip-is-built-for-fast-inference-at-scale-benchmarks-show\u002F)。\n\n## 700 瓦背后的取舍\n\n这颗 ASIC 标称 700W TDP,实测持续功率「at or below 550W」,而对比系统 GB200 \u002F GB300 分别是 1,200W \u002F 1,400W。功耗低一半,单芯片 TDP 砍到 Blackwell 的二分之一,但单用户吞吐却高出 2–4 倍——这意味着 OpenAI 用系统级协同换性能:数据搬运少、网络域大、KV cache 显式放置。\n\nRichard Ho 自己划了底线:这些数字都是「单轮 8k context,InferenceX」跑出来的,**没有 AgentX、没有多轮、没有 prefix-cache-heavy 场景**。也就是说,今天 Jalapeño 证明的是「低成本 LLM 推理」很能打,**而真正赚钱的 agent workload 还没验证**。这也是为什么 Hot Chip 2026 上,SemiAnalysis 内部自己的复盘是「Jalapeño vs Blackwell 的对比不公平也不完整」—— Nvidia 的下一代 Rubin 已经用同样 HBM4 内存出货,功耗 900–1,150W,反而成了「Jalapeño 真正要打的对手」。\n\n## AI 帮 AI 写 kernel:更值得追踪的故事\n\n硬件本身是表层。真正让从业者坐直身子的是另一段披露:\n\n> 「Using Codex with GPT-Astra, the team brought three open-weight models that were not part of Jalapeño's original production plan to high performance within two months. For selected GPT-OSS attention and mixture-of-experts blocks, **AI-generated implementations ran 1.5 to 1.8 times faster than the existing human-expert-written implementations**。」\n\n翻译一下:Jalapeño 从硅回来到上线能跑生产模型,只用了 **不到九个月**;而其中一部分 GPT-OSS attention 和 MoE blocks 的 kernel,**是 Codex 加 GPT-Astra 自动写出来的**,在选定的 block 上跑得比人类专家手写的还快 1.5–1.8 倍。OpenAI 给出的不是一句 PR 话术,而是一个**可以量化的证据链**——AI 不仅能写代码,还能写出比资深 GPU kernel 工程师更快的 kernel。\n\n如果这条曲线继续放大,「**十年 CUDA 生态积累的护城河**」就要被重新计价。OpenAI 自研芯片单点收益有限,但「模型→硬件」的双向耦合一旦成立,**任何新模型架构都能在九个月内跑出硅级加速**,这是 Nvidia 给不了的迭代速度。CUDA 的护城河本质是「人写 kernel」的规模优势,不是「代码本身」;一旦写 kernel 这件事被模型接管,生态累积的复用价值就会快速折旧。\n\n## 那 Nvidia 这下慌不慌?\n\nOpenAI 自己也没把话说死:「我们会继续大量部署 NVIDIA 和其他合作伙伴的加速器」。Jalapeño 不是用来替代 Nvidia,而是给 OpenAI 自己 API 的 cost-per-token 多一层下压空间——按他们估算,如果推理成本真的降 50%,那高频 agent 工作流就不再卡死在「贵到用不起」这条线上。\n\n半年前 SemiAnalysis 自己评价过:「Jalapeño 跟 Blackwell 的对比是不公平也有些不完整」。如果把对照对象换成已出货的 Nvidia Vera Rubin(同样的 HBM4 内存、900–1,150W 功耗),**Jalapeño 在 TCO 上跟 Rubin 大致打平,而不是碾压**。CUDA 生态仍然值钱,只是「唯一选项」的窗口正在关闭——TPU、Trainium、Jalapeño 之后,头部 AI 实验室不再「只能买 Nvidia」,议价权开始转移。\n\n## 给开发者意味着什么\n\n现在调 OpenAI 的 API,代码一行都不用改。Jalapeño 是内部基础设施,不是给你插一块的硬件卡。**真正的红利要等到 2027 年,API 价格或 rate-limit 上看得到的那次调整**。在那之前,值得追踪的信号只有三个:\n\n1. Gen 2、Gen 3 的演进曲线(OpenAI 已经公开承认在做了);\n2. AgentX 长上下文、多轮、prefix-cache 场景下,Jalapeño 是否还能稳坐 Pareto 前沿;\n3. 「Codex 写 kernel 比人快 1.5 倍」是否从选定的 blocks 扩散到整张芯片的代码库。\n\n如果三件事同时发生,2027 年的 LLM 推理市场就不是「Nvidia vs 自研硅」这么简单的二元对立,而是变成「**模型-硬件紧耦合的全栈玩家**」对「**通用芯片 + 老牌编译器护城河**」的代际较量。\n\n参考来源:\n- [Jalapeño's first results - OpenAI Engineering](https:\u002F\u002Fopenai.com\u002Findex\u002Fjalapeno-first-results\u002F)\n- [OpenAI's Jalapeño chip is built for fast inference at scale - TechCrunch](https:\u002F\u002Ftechcrunch.com\u002F2026\u002F08\u002F25\u002Fopenais-jalapeno-chip-is-built-for-fast-inference-at-scale-benchmarks-show\u002F)\n- [OpenAI and Broadcom unveil LLM-optimized inference chip - OpenAI](https:\u002F\u002Fopenai.com\u002Findex\u002Fopenai-broadcom-jalapeno-inference-chip\u002F)","https:\u002F\u002Ftechcrunch.com\u002F2026\u002F08\u002F25\u002Fopenais-jalapeno-chip-is-built-for-fast-inference-at-scale-benchmarks-show\u002F","226bcb3d-18b8-4bb0-a999-4e82ec13f5fd",[11,15,18,21,24],{"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},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":25,"name":26,"slug":26,"description":14,"color":14},"42e59a88-7795-47dc-a334-ef1e72c24347","openai",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"d18cc3ab-d298-4643-a364-011528d075f2","en","OpenAI Jalapeño Hot Chip Bench: 1.9× per Watt, 1s Latency","Hot Chips 2026: OpenAI disclosed Jalapeño benchmarks on SemiAnalysis InferenceX. Vs Nvidia Blackwell: peak throughput per watt 1.5-1.9x, end-to-end latency 1.7-3.6x lower, per-user tokens\u002Fs reaches 1,459 on GPT-OSS vs GB200 535.","OpenAI's self-designed inference silicon has been rumored since last year, but on August 25 at Hot Chips 2026, Richard Ho's team finally put numbers on the board. This is OpenAI's first real attempt to answer two questions: how much faster is a custom ASIC than Nvidia Blackwell, and whether a nine-month tape-out miracle is held together by AI writing its own kernels.\n\n## The Real Numbers in Three Charts\n\nOpenAI did not pick its strongest internal model. Instead, it ran three public models through SemiAnalysis's InferenceX benchmark so anyone can reproduce the results:\n\n- **GPT-OSS 120B**: peak throughput 85,448 vs 44,960 mixed tokens\u002Fs per kW (vs GB200), a **1.9x** lead; end-to-end latency dropped from 1.80 s to 1.03 s.\n- **DeepSeek R1 670B (MXFP4)**: 19,641 vs 11,781 mixed\u002FkW, **1.7x**; latency 1.65 s vs 5.99 s.\n- **Kimi K2.5 1T (MXFP4)**: 18,195 vs 11,862 mixed\u002FkW, **1.5x**; latency 1.56 s vs 5.31 s.\n\nThe per-user story is even sharper. On DeepSeek R1, Jalapeño pushes **700 tokens\u002Fs** per user vs 169 on the comparison system. On Kimi K2.5: **694 vs 182**. On GPT-OSS: **1,459 vs 535**. In other words, the \"is it snappy\" gap reaches **over 4x** at the user level.\n\nSource: [OpenAI's official disclosure page](https:\u002F\u002Fopenai.com\u002Findex\u002Fjalapeno-first-results\u002F) and [TechCrunch's report](https:\u002F\u002Ftechcrunch.com\u002F2026\u002F08\u002F25\u002Fopenais-jalapeno-chip-is-built-for-fast-inference-at-scale-benchmarks-show\u002F).\n\n## The 700-Watt Tradeoff\n\nThe ASIC carries a 700 W TDP rating, with measured sustained power \"at or below 550 W,\" while the comparison systems (GB200 \u002F GB300) sit at 1,200 W \u002F 1,400 W. Power is halved, single-chip TDP is roughly half of Blackwell, but per-user throughput is 2–4x higher — meaning OpenAI is trading system-level co-design for performance: less data movement, larger network domains, and explicit placement of the KV cache.\n\nRichard Ho drew a clear line himself: all numbers are single-turn, 8k context, InferenceX. **No AgentX, no multi-turn, no prefix-cache-heavy workloads**. Jalapeño today proves \"cheap LLM inference\" can dominate. The agent workloads that actually pay the bills are still unverified. That is also why SemiAnalysis's own postmortem at Hot Chips 2026 called the Jalapeño vs Blackwell comparison \"unfair and incomplete\" — Nvidia's next-generation Rubin is already shipping with the same HBM4 memory and 900–1,150 W power envelope, which is the real opponent Jalapeño will have to face.\n\n## The More Important Story: AI Writing Kernels\n\nThe hardware is surface-level. The disclosure that made practitioners sit up:\n\n> \"Using Codex with GPT-Astra, the team brought three open-weight models that were not part of Jalapeño's original production plan to high performance within two months. For selected GPT-OSS attention and mixture-of-experts blocks, **AI-generated implementations ran 1.5 to 1.8 times faster than the existing human-expert-written implementations**.\"\n\nTranslate this: from silicon return to production-ready models, Jalapeño took **less than nine months**; part of the GPT-OSS attention and MoE block kernels were **auto-written by Codex plus GPT-Astra**, and on selected blocks ran **1.5–1.8x faster than human-expert kernels**. OpenAI is not offering PR copy, it is offering a **quantifiable evidence chain** — AI can not only write code, it can write code that beats veteran GPU kernel engineers.\n\nIf this curve scales, the **decade of accumulated CUDA moat** has to be repriced. OpenAI's single-chip win is bounded; what matters is the bidirectional coupling between \"models\" and \"hardware.\" Once it exists, **any new model architecture can ship at silicon speed within nine months**, an iteration cadence Nvidia cannot match. The CUDA moat was always \"humans writing kernels at scale,\" not \"the code itself.\" Once the writing is done by models, the accumulated ecosystem amortization disappears fast.\n\n## Does Nvidia Panic?\n\nOpenAI itself hedged: \"we will continue to deploy NVIDIA and other partners' accelerators at scale.\" Jalapeño is not designed to replace Nvidia; it gives OpenAI's own API another lever to push down cost-per-token. By their estimate, if inference cost drops 50%, high-frequency agent workflows no longer hit the wall of \"too expensive to use.\"\n\nSemiAnalysis itself noted earlier that the Jalapeño vs Blackwell comparison was \"unfair and incomplete.\" Against shipping Nvidia Vera Rubin (same HBM4 memory, 900–1,150 W), **Jalapeño lands at rough parity on TCO, not a wipeout**. The CUDA ecosystem still has value; the \"only choice\" window is just closing. After TPU, Trainium, and Jalapeño, frontier labs are no longer locked into \"Nvidia only,\" and pricing power starts to shift.\n\n## What This Means for Developers\n\nIf you call the OpenAI API today, you change no code. Jalapeño is internal infrastructure, not a card you can plug into a server. **The real payoff arrives in 2027**, when the next API pricing or rate-limit adjustment reflects the cost structure. Until then, three signals are worth tracking:\n\n1. The evolution curve of Gen 2 and Gen 3 (OpenAI has publicly confirmed both are in development);\n2. Whether Jalapeño holds the Pareto frontier on AgentX long-context, multi-turn, and prefix-cache-heavy workloads;\n3. Whether \"Codex writes kernels 1.5x faster than humans\" generalizes from selected blocks to the full chip code base.\n\nIf all three happen at once, 2027's LLM inference market will not be a binary \"Nvidia vs custom silicon\" fight — it will be a generational contest between **vertically integrated model-hardware stacks** and **general chips plus legacy compiler moats**.\n\nReference sources:\n- [Jalapeño's first results - OpenAI Engineering](https:\u002F\u002Fopenai.com\u002Findex\u002Fjalapeno-first-results\u002F)\n- [OpenAI's Jalapeño chip is built for fast inference at scale - TechCrunch](https:\u002F\u002Ftechcrunch.com\u002F2026\u002F08\u002F25\u002Fopenais-jalapeno-chip-is-built-for-fast-inference-at-scale-benchmarks-show\u002F)\n- [OpenAI and Broadcom unveil LLM-optimized inference chip - OpenAI](https:\u002F\u002Fopenai.com\u002Findex\u002Fopenai-broadcom-jalapeno-inference-chip\u002F)","openai-jalapeno-hot-chips-benchmark-2026","2026-08-26T02:00:00Z","2026-08-26T07:10:00.384382Z","2026-08-26T07:10:00.384390Z",true,"agent",22,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"54d25a48-2524-41d5-8043-da7d3d88c7cd","OpenAI 联合 Broadcom 推出 Jalapeño：专为 LLM 推理自研，9 个月从设计到流片","openai-broadcom-jalapeno-llm-inference-chip","2026-06-24T14:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"68072ee1-fc37-4064-ab18-09550ae72d1b","GLM-5.3-Flash 把 320B MoE 跑在国产芯片上:Flash 价位和 $0.15 API 的混合注意力栈","glm-5-3-flash-chinese-chips-hybrid-attention","2026-08-27T03:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"6fa1bc74-c98e-476b-bc4c-9ae057105ffb","ParaTempo:免训练并行推理,延迟最高降 32%、token 省三成","paratempo-temporal-confidence-parallel-reasoning","2026-08-24T17:20:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"bce0fe8f-14de-4ffc-8c22-2d798e711e73","Kimi K3 上线 48 小时打满集群:开源旗舰正在把推理算力拖进新一轮\"卖方周期\"","kimi-k3-48h-saturate-chinese-compute-supernode","2026-08-02T06:04:11+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"de219584-58fc-45ad-91ea-0049a5cbcf10","OpenAI 开源 Codex Security CLI:把 AI 安全检测塞进每个 PR","openai-codex-security-cli-opensource","2026-07-29T10:30:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"b0183d10-bcfd-44ed-a178-a2c813f10b69","国家超算互联网AI社区上线Kimi K3:2.8万亿参数MoE一键调用,开源大模型有了国产算力底座","kimi-k3-cnsc-internet-launch","2026-07-28T09:30:00+00:00"]