盗墓笔记

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

一 |     (ECNS) -- More Chinese provinces are rolling out paid‑leave policies aimed at turning time off into consumer spending, with Henan urging officials to lead by example.    Henan issued a notice Wednesday requiring employers to ensure all eligible employees take their full annual leave, while encouraging scenic spots and rural guesthouses to design packages around music festivals, farming experiences, Yellow River leisure, ancient towns and folk activities.    At least six other provincial‑level regions — Yunnan, Guizhou, Shaanxi, Shanxi, Anhui and Tianjin — have introduced similar measures this year. The State Council's work report called for staggered paid leave for the first time, and the latest five‑year plan reaffirmed the system.    According to Song Xiangqing, vice president of the China Society of Business Economics, the longstanding over-reliance on statutory holidays has caused a tidal pattern in tourism, with overcrowding during peak seasons and underuse in off-peak periods.    Staggered paid leave policy aligns closely with the broader national drive to boost domestic demand and stimulate consumption, and is likely to turn leisure time into a steady source of spending momentum, Song said.                        。    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。

二 |          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。

三 |          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。

四 |          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。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。

五 | 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。         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。    

Current article:http://www.mengruanpianchengji.cyou/lhqi0kz/y08.ppt

Published on:18:48:28