时间: 00:47:04 来源: 星际特工 作者: 我的网站 点击: 55330
我们的少年时代

向幸福出发

30倍吞吐碾压!英伟达Vera Rubin首测出炉:Agent工作负载下GB300被彻底超越_我的网站

舞出我人生

A |     8月25日消息,英伟达近日在Hot Chips 2026大会上首次公布了下一代旗舰机柜Vera Rubin NVL72的片上实测数据。

B |     Jair Bolsonaro has tried to snatch social media influencer Wilker Leao’s mobile phone after the YouTuber confronted the Brazilian president, calling him a “coward” and a “vagabond”, among other taunts.,The incident took place as Bolsonaro spoke to supporters and posed for selfies outside his official residence in the capital Brasilia on Thursday.,A video released by the Brazilian news station Globo showed the Brazilian president attempting to grab Leao’s shirt and forearm and his bodyguards then pulling the YouTuber aside. A few moments on, the social media influencer was allowed by security to return and speak to Bolsonaro for about five minutes.,This followed the beginning of the presidential campaign in Brazil on Tuesday, with Bolsonaro still trailing his left-wing rival, former President Luiz Inacio Lula da Silva in all polls ahead of the October 2nd first round vote.。          此次实测直接调用开源模型DeepSeek V4 Pro(1.6万亿参数),运行真实智能体编码任务,结果令人震惊。         在SemiAnalysis旗下AgentX基准测试中,Vera Rubin NVL72的每兆瓦吞吐量,比当前主力机皇GB300 NVL72最高飙升30倍,生产每百万Token的成本最高暴降35倍。     作为对比,GB300的每兆瓦吞吐量相较上代H200也已提升15倍,Vera Rubin是在这个基础上再拔高30倍。

C |          英伟达借此宣告旧基准作废:传统8K或1K固定序列测试已无法衡量智能体,必须回放含上下文增长与工具调用的真实工作流。

D | 数据显示,一次Agent任务消耗的Token是普通聊天的15倍,长上下文成为关键瓶颈。    性能跃升来自极致协同设计。NVFP4量化将权重压至4位精度,第六代NVLink配合大模型MoE提供比以太网快10倍、延迟低3倍的互联,让DeepSeek这类专家模型在72颗GPU间行云流水调度。

E |          同场还有两枚新芯量产:一是专为低延迟定制的Groq 3 LPX让Gemma 4 31B跑出每秒3400 Token;二是面向智能体打造的Vera CPU,含88个自研Olympus核心、内存带宽1.2TB/s。         值得一提的是Vera CPU已被马斯克SpaceXAI连夜装机,准备部署上太空。

F |     

         

     【本文结束】如需转载请务必注明出处:      责任编辑:红茶     文章内容举报     

Current article:http://724xs.runtunpaixufenjuduishuang.pics/list_xht/6jxpg.html

Published on:11:21:45


本文标签: 全职妈妈 源代码 大秦帝国之纵横

回到顶部