中国式相亲

英伟达详解 Vera CPU：专为智能体 AI 构建，88 个Olympus 核心_我的网站

A |     IT之家 7 月 22 日消息，英伟达昨日（7 月 21 日）发布博文，从技术角度详细介绍了 Vera CPU，专为智能体 AI（Agentic AI）场景设计。    

    A service area along the nearly completed Pinglu Canal in Qinzhou, South China's Guangxi Zhuang Autonomous Region, on June 3, 2026. Photo: VCG
The Pinglu Canal in South China's Guangxi Zhuang Autonomous Region started live-ship navigation trial on Wednesday. It is the first full-channel, all-vessel-type live-ship trial on the canal and the last test before the waterway fully opens to navigation in September, the CCTV News reported. 
Four vessels of the 1,000-ton, 3,000-ton and 5,000-ton classes were deployed for the trial. Moving in phases along the canal's main channel, they will verify navigation under different operating conditions and test the performance of navigation aids across the entire route, comprehensively checking the safety and stability of the canal's channel, ship-locks, anchorages and service areas to lay a solid foundation for the canal's opening, said the report. 
The Pinglu Canal is a backbone project of the New International Land-Sea Trade Corridor and the first canal linking inland rivers to the sea to be coordinated at the national level since the founding of the People's Republic of China. Overall construction progress has passed 98 percent, the report noted. 
Notably, the canal's planned opening month of September coincides with the annual China-ASEAN Expo in Nanning, Guangxi.
The canal stretches 134.2 kilometers from the Pingtangjiangkou section of the Xijin reservoir area in Nanning, to the Beibu Gulf via Qinzhou city. It can handle 5,000-ton vessels.
Three navigation hubs — Madao, Qishi and Qingnian — are located from upstream to downstream, equipped with twin single-step 5,000-ton ship-locks with a designed annual one-way capacity of 89 million tons. Construction began in August 2022, with a total budgeted cost of 72.719 billion yuan ($10.78 billion), according to Nanning Evening News.
Once in operation, the canal will let ships from the middle and upper reaches of the Xijiang River sail directly to the Beibu Gulf. The canal will cut roughly 560 kilometers from the current inland waterway route to the sea via Guangzhou Port, shortening delivery time and lowering logistics costs, said Xinhua.
The first full‑route patrol of the Pinglu Canal was completed on August 8. During the patrol, maritime authorities conducted on‑site inspections of key navigation conditions, including channel dimensions, curved sections of the waterway, navigational clearance beneath bridges, ship‑lock approach channels and more, the Guangxi Daily reported. 
Global Times
。         英伟达表示智能体 AI 将更多的关键执行路径转移到了 CPU 上。智能体（Agents）在沙箱中运行代码、调用工具、检索上下文、与数据库交互以及分析结果，然后再将信息返回给模型。         IT之家援引博文介绍，智能体 AI 工作负载主要取决于以下几点：          即使插槽完全负载，也需要强大的单线程性能。    每个核心都有足够的内存带宽，以持续为许多活跃的执行上下文提供数据。    并发下的延迟具有可预测性，以便智能体能够稳定完成预定步骤。    高效处理不规则控制流、长依赖链和指针密集型数据结构。         英伟达在博文中强调，Vera CPU 围绕智能体 AI 负载设计，重点提升满负载条件下的单线程表现，其设计核心是 Olympus 核心。

B |          英伟达强调 Olympus 核心采用前端、核心中段、执行引擎和缓存子系统四部分设计。前端集成 10 宽解码引擎，并支持每周期最多处理 2 个已采用分支。其分支预测子系统包含神经分支预测器，用于提升复杂分支模式下的预测准确性。         该公司表示，Vera CPU 共配备 88 个 Olympus 核心和 176 个 SMT 线程，目标是在单线程高负载与更高线程密度两类场景间切换，并降低线程间资源争用。         在片上互连与缓存部分，Vera CPU 搭载 NVIDIA 可扩展相干互连，即 NVIDIA Scalable Coherency Fabric（SCF）。         NVIDIA 称，这一互连提供最高 3.4 TB/s 双向分区带宽，并在全核心范围内集成 164 MB 统一 L3 缓存，连接核心、共享缓存、内存控制器、I/O 与 NVLink-C2C 接口。

C |          内存方面，Vera CPU 采用 SOCAMM2 LPDDR5X 内存。

D | NVIDIA 称，该子系统聚合带宽最高 1.2 TB/s，按单核计算最高 14 GB/s，面向高并发 AI、分析、图计算与强化学习负载供数。         扩展与 I/O 方面，NVIDIA 称 Vera CPU 支持双路扩展，通过第二代 NVLink-C2C 构建一致性双路平台。         双路配置下，平台提供 176 条 PCIe 6.4 通道，并支持 CXL 3.1。保密计算方面，Vera CPU 基于 Arm CCA / RME，支持按虚拟机分配加密密钥，并使用经认证的 C2C 加密保护一致性链路数据。

Current article:http://www.xiongpeiduanlouduoennongleng.cyou/hxd35/20260826/904667.htm

Published on:05:55:50