当前位置:主页 > 国内 >

康熙王朝

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

大明王朝1566

A |     Colombo, Oct 17 (UNI) The Sri Lankan government should withdraw a draft law that will give the authorities broad powers to detain people in military-run “rehabilitation” centers, placing them at great risk of abuse, Human Rights Watch (HRW) has said.
The HRW said the Bureau of Rehabilitation Bill, submitted to Parliament on September 23, would allow the compulsory detention in centers of “drug dependent persons, ex-combatants, members of violent extremist groups and any other group of persons”, the Daily Mirror reported on Monday.
“The Bureau of Rehabilitation Bill would establish a new administrative structure controlled by the Defense Ministry to operate 'rehabilitation' centers staffed by military personnel.
"The proposed law, which human rights advocates have already challenged in the Supreme Court, does not describe the basis for being sent for 'rehabilitation' but other recent government policies provide vague and arbitrary powers to forcibly 'rehabilitate' people who have not been convicted of any crime,” the HRW said.
“The Sri Lankan government’s proposed ‘rehabilitation’ efforts appear to be nothing more than a new form of abusive detention without charge,” said Meenakshi Ganguly, South Asia director at Human Rights Watch. “The Rehabilitation Bill would open the door widely to more torture, mistreatment, and endless detention.”
The Rehabilitation Bill is the latest measure in a long history of laws, such as the Prevention of Terrorism Act (PTA), that authorize arbitrary detention and torture in Sri Lanka. The law could be used to target minority communities or anti-government protesters.
UNI MR。     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。

B | 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%。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。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。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。

C | 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。

D | 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。

E | 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:。

Current article:http://50178.ruwentaqintiaozhupapaigai.shop/list_f0wyzv/d1r.html

Published on:16:54:58


相关新闻

最后更新

热门新闻