主页 > 国内 >

那一场呼啸而过的青春

OpenAI 首款自研芯片 Jalapeño 性能首秀:DeepSeek R1 每瓦 AI 吞吐量是 GB300 的 1.7 倍_我的网站

罗马假日

A |      8 月 26 日消息,OpenAI 昨日(8 月 25 日)发布博文,展示其首款自研 Jalapeño 芯片,并表示无论是单用户 token 处理量,还是每千瓦吞吐量,都超过了目前市面上的顶尖推理处理器。在官方博文中,OpenAI 使用半导体研究公司 SemiAnalysis 开发的公共基准测试工具 InferenceX,通过 GPT-OSS 120B、DeepSeek R1 670B 和 Kimi K2.5 1T 三款模型测试。测试显示对比市面上领先的 AI 系统(GB200 和 GB300),每瓦能够完成的 AI 吞吐量是对比系统的 1.5 至 1.9 倍,端到端延迟则约为对比系统的 28% 至 59%。    The senator told the Greek news agency that the United States would sell military equipment to those who share its values, "not just for financial gain," and listed a number of problems on which Ankara should change its behavior and attitude. Among them are the issue of acquiring the Russia's S-400 system, flights over the Greek islands, Turkey's close relations with Russia, and human rights violations, including the imprisonment of journalists and lawyers.,Menendez added that the list of outstanding issues is long, and they must be resolved in order to achieve the sale and modernization of the F-16 jets, but he had not yet seen the readiness of Turkish President Recep Tayyip Erdogan to take any of these steps, the news agency reported.,Menendez is considered a person who "holds the keys to the F-16," since he has the power to veto the sale of the US weapons systems as the chairman of the Committee on Foreign Relations, the agency said.,Ankara Hopes for Success in F-16 Deal With US, Turkish Defence Minister Says10 May, 14:41 GMT,Turkey plans to purchase 40 F-16 fighter jets from the US, and modernize another 80 which it already has in service. US President Joe Biden said that he expects to obtain a congressional approval for the sale of the jets to Turkey, but a group of the US congressmen signed a statement objecting to the sale.。

B | 混合 Token 吞吐量与现有 AI 加速器的最快解码速度对比3 款模型峰值单用户解码吞吐量各平台最佳运行点的混合 Token 峰值吞吐量OpenAI 表示,Jalapeno 在运行 GPT-OSS 120B 时,每千瓦的峰值吞吐量比 GB200 系统高出约 1.9 倍。

C | 在 DeepSeek R1 上,它的每瓦性能比 GB300 高出约 1.7 倍;在 Kimi K2.5 上,它的每瓦性能比 GB300 高出约 1.5 倍。

D | Jalapeno 的额定功率为 700 瓦,但在测试工作负载期间,其持续功耗保持在 550 瓦或以下。附上相关图片如下:延伸阅读Jalapeño 是 OpenAI 首款自研 AI 推理芯片,由 OpenAI 与博通(Broadcom)合作开发,台积电采用 3nm 制程代工制造,于 2026 年 6 月 24 日首次对外展示。该芯片是一颗 ASIC(专用集成电路),采用 systolic array(脉动阵列)架构,配备 HBM 高带宽内存,专为大语言模型推理场景设计,不用于模型训练。芯片及配套服务器系统均不会对外销售,仅供 OpenAI 内部使用,旨在降低自身 GPU 采购和算力运营成本。OpenAI 硬件负责人 Richard Ho 领导了该芯片的研发,他此前在 Google 工作近九年,是 Cloud TPU 项目的核心工程师。Jalapeño 从架构设计到完成流片仅耗时约 9 个月,OpenAI 称这是高性能先进半导体领域有史以来最快的 ASIC 开发周期之一。OpenAI 自身的 AI 模型也深度参与了芯片设计流程,帮助加速设计优化。目前第二代 Jalapeño 芯片已进入后期开发阶段,预计未来数月内完成流片;第三代芯片也已启动概念设计。OpenAI 计划于 2026 年年底前小规模部署 Jalapeño,2027 年逐步扩大部署规模。需要注意的是,Jalapeño 此次测试对比的对象为英伟达 GB200 和 GB300 系统,尚未与英伟达最新一代 Vera Rubin 芯片进行对比测试。OpenAI 表示英伟达仍是其重要合作伙伴,未来将持续采用多供应商策略以满足庞大的算力需求。相关阅读:。

E |

Current article:http://p15al.guaxiazaopinduan.cfd/d0f/gl9sfmd.html

Published on:00:00:00


相关文章
推荐图文
最热文章