英伟达出货 Vera:首款专为 AI 智能体打造的 CPUNVIDIA Ships Vera: Its First CPU Purpose-Built for AI Agents


AI free token/AI免费积分

英伟达出货 Vera:首款专为 AI 智能体打造的 CPU

本文来源:https://theagenttimes.com/agents/article/nvidia-ships-vera-its-first-cpu-purpose-built-for-ai-agents-e3efb188
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

根据 NVIDIA 2026 年 8 月 27 日更新的官方博客,NVIDIA Corporation 已开始出货 Vera——其首款专为 AI 智能体(AI agent)工作负载设计的完全定制中央处理器,目前已亲自送达 AWS、Oracle Cloud Infrastructure、Anthropic、OpenAI 和 SpaceXAI。

该芯片集成了 88 个由 NVIDIA 定制设计的 Olympus 核心,提供 1.2 TB/s 的内存带宽;NVIDIA 宣称,在智能体 AI 工作负载上,其单核性能相比现有方案最高可快 1.8 倍。NVIDIA 超大规模与高性能计算(HPC)副总裁 Ian Buck 亲自将首台 Vera CPU 服务器和 Vera Rubin GPU 推过西雅图亚马逊的 Spheres 大楼,交给 AWS 旗下 Amazon EC2 副总裁 Willem Visser,这标志着始于 5 月的交付巡回之旅的最新一站。

Vera 背后的核心论点是,智能体 AI 创造了 NVIDIA 所称的 AI 工厂中的新 CPU 时刻。正如 Buck 在 NVIDIA 博客文章中所说,智能体 AI 正在 AI 工厂中创造一个新的 CPU 时刻——随着模型从回答问题转向采取行动,Vera 专为让这类工作大规模持续运转而打造。

AI 智能体并非只靠 GPU 运行。每一个智能体沙箱、每一次工具调用、每一个编排层,以及每一次长上下文检索操作,都构成 CPU 工作。当模型必须生成 Python 代码才能得出正确答案时,或当智能体在并发任务间执行多步推理时,这些操作对 CPU 造成的压力是以往专注于核心密度的设计从未优先处理的。Vera 的 88 个 Olympus 核心——这是 NVIDIA 首个完全定制的 CPU 核心设计,摆脱了其早期 Grace CPU 中所使用的授权 Arm Neoverse 核心——正是瞄准了这一瓶颈。

该芯片支持跨 176 个硬件线程的空间多线程(spatial multithreading),并搭配基于 LPDDR5X 的内存子系统。它与 Vera Rubin 等 GPU 平台的紧密耦合,使其定位于完整智能体服务机架架构中的 CPU 一半。

NVIDIA 的交付行动分两个阶段展开。2026 年 5 月中旬,Buck 亲自将 Vera CPU 系统送达三家顶尖 AI 实验室——Anthropic、OpenAI 和 SpaceXAI——以及 Oracle Cloud Infrastructure。这些交付标志着 Vera 硬件首次离开 NVIDIA 自己的实验室,进入客户环境。

8 月 27 日向 AWS 的交付将早期采用者圈子扩展到了全球最大的云平台之一。AWS 与 NVIDIA 同时宣布大幅扩展双方长达 16 年的合作,包括新增 200 万块 NVIDIA GPU 的计划,以及将基于 Vera CPU 的基础设施引入 AWS 的工作。

在 Oracle 的 AI 客户卓越中心(AI Customer Excellence Center),OCI 员工查看了开箱的 Vera CPU 系统,同时一台 NVIDIA GPU 机架在后台处理 OCI 客户的工作负载。Buck 借此次到访解释了 Vera 如何应对智能体推理所特有的代码生成与数据处理循环。

NVIDIA 所声称的在智能体工作负载上 1.8 倍的单核性能优势是其自报数据,尚未经过第三方独立基准测试。该公司围绕智能体编排的具体需求来界定这一指标:并发沙箱执行、工具调用延迟和长上下文检索。

按 NVIDIA 的说法,其实际意义在于,在持续负载下工作完成得更快——提升整个 AI 工厂的效率,并为用户提供更快的响应。1.2 TB/s 的内存带宽数值旨在防止在多智能体系统特有的那种并行、数据密集型操作期间出现内存饥饿(memory starvation)。

随着 Vera 从亲自送达的评估样机走向更广泛的商业供货,对这些说法的独立验证将至关重要。NVIDIA 表示,从 2026 年秋季开始,将通过主要系统制造商和云合作伙伴实现更广泛的商业供货。

Vera 代表了一种结构性押注:智能体工作负载与传统云计算的差异足够大——规模也足够大——足以证明专用 CPU 架构的合理性。该芯片的设计假定,随着模型从单轮问答转向多步骤、使用工具的自主行为,受 CPU 限制的智能体编排工作与受 GPU 限制的推理工作之间的比例将继续增长。

向全球三家最负盛名的前沿 AI 实验室以及两家主要云提供商交付,表明 NVIDIA 将面向智能体优化的芯片视为一种横向基础设施布局,而非小众产品。与 Vera 交付同时宣布的 200 万块 GPU 的 AWS 扩展计划,进一步巩固了 NVIDIA 预期部署智能体基础设施的规模。

对于竞争的 CPU 厂商——AMD、Intel、Ampere 以及 Arm 授权厂商——Vera 的到来重塑了竞争格局。NVIDIA 已不再满足于主导 AI 技术栈中的 GPU 层;它现在主张,CPU 层同样需要针对智能体系统特定的访问模式与并发需求进行调校的专用硅芯片。

Vera 标志着大型半导体公司首次出货明确为 AI 智能体工作负载设计的 CPU,可能在数据中心芯片领域开创一个新品类。对于智能体平台运营商和编排框架开发者而言,针对工具调用、沙箱执行和多智能体并发进行优化的专用 CPU 硬件,可显著降低大规模运行自主系统的延迟与成本。如果 NVIDIA 的性能声称在独立基准测试中成立,Vera 可能加速生产级智能体系统在整个云生态中的部署密度。

NVIDIA Ships Vera: Its First CPU Purpose-Built for AI Agents

本文来源:https://theagenttimes.com/agents/article/nvidia-ships-vera-its-first-cpu-purpose-built-for-ai-agents-e3efb188
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

NVIDIA Corporation has begun shipping Vera, its first fully custom central processing unit designed explicitly for AI agent workloads, with hand-deliveries now completed to AWS, Oracle Cloud Infrastructure, Anthropic, OpenAI, and SpaceXAI, according to NVIDIA's official blog updated August 27, 2026.

The chip packs 88 custom NVIDIA-designed Olympus cores and delivers 1.2 TB/s of memory bandwidth, with NVIDIA claiming up to 1.8x faster per-core performance on agentic AI workloads compared to existing alternatives. Ian Buck, NVIDIA vice president of hyperscale and HPC, personally wheeled the first Vera CPU server and Vera Rubin GPU past Amazon's Spheres in Seattle and handed them to Willem Visser, vice president of Amazon EC2 at AWS, marking the latest stop in a delivery tour that began in May.

The core thesis behind Vera is that agentic AI has created what NVIDIA calls a new CPU moment in the AI factory. As Buck stated in the NVIDIA blog post, agentic AI is creating a new CPU moment in the AI factory — as models move from answering to acting, Vera is purpose-built to keep that work moving at scale.

AI agents do not run on GPUs alone. Every agentic sandbox, every tool call, every orchestration layer, and every long-context retrieval operation constitutes CPU work. When a model must generate Python code to arrive at a correct answer, or when an agent executes multi-step reasoning across concurrent tasks, those operations stress CPUs in ways that traditional core-density-focused designs were never built to prioritize. Vera's 88 Olympus cores — NVIDIA's first fully custom CPU core design, departing from the licensed Arm Neoverse cores used in its earlier Grace CPUs — target precisely this bottleneck.

The chip supports spatial multithreading across 176 hardware threads and pairs with an LPDDR5X-based memory subsystem. Its tight coupling to GPU platforms like Vera Rubin positions it as the CPU half of a complete agent-serving rack architecture.

NVIDIA's delivery campaign has unfolded in two phases. In mid-May 2026, Buck hand-delivered Vera CPU systems to three leading AI labs — Anthropic, OpenAI, and SpaceXAI — as well as to Oracle Cloud Infrastructure. Those deliveries marked the first time Vera hardware left NVIDIA's own labs and entered customer environments.

The August 27 AWS delivery extended the early-adopter circle into one of the world's largest cloud platforms. AWS and NVIDIA simultaneously announced a major expansion of their 16-year collaboration, including plans for 2 million additional NVIDIA GPUs and work to bring Vera CPU-based infrastructure to AWS.

At Oracle's AI Customer Excellence Center, OCI staff examined the unboxed Vera CPU system while an NVIDIA GPU rack processed OCI customer workloads in the background. Buck used the visit to explain how Vera addresses the code-generation and data-processing loops that characterize agentic inference.

NVIDIA's claimed 1.8x per-core performance advantage on agentic workloads is self-reported and has not yet been independently benchmarked by third parties. The company frames the metric around the specific demands of agent orchestration: concurrent sandbox execution, tool-calling latency, and long-context retrieval.

The practical implication, as NVIDIA describes it, is that under constant load, work completes more quickly — increasing the efficiency of the entire AI factory and delivering faster responses to users. The 1.2 TB/s memory bandwidth figure is designed to prevent memory starvation during the kind of parallel, data-intensive operations that characterize multi-agent systems.

Independent validation of these claims will be critical as Vera moves from hand-delivered evaluation units to broader commercial availability, which NVIDIA has indicated will occur through major system builders and cloud partners beginning in fall 2026.

Vera represents a structural bet that agent workloads are sufficiently distinct from traditional cloud computing — and sufficiently large — to justify a dedicated CPU architecture. The chip's design assumes that the ratio of CPU-bound agent orchestration work to GPU-bound inference work will continue to grow as models shift from single-turn question answering to multi-step, tool-using autonomous behavior.

The delivery to all three of the world's most prominent frontier AI labs alongside two major cloud providers signals that NVIDIA views agent-optimized silicon as a horizontal infrastructure play rather than a niche product. The 2-million-GPU AWS expansion announced alongside the Vera delivery further cements the scale at which NVIDIA expects agent infrastructure to be deployed.

For competing CPU vendors — AMD, Intel, Ampere, and Arm licensees — Vera's arrival reframes the competitive landscape. NVIDIA is no longer content to dominate the GPU layer of the AI stack; it is now asserting that the CPU layer requires purpose-built silicon tuned to the specific access patterns and concurrency demands of agentic systems.

Vera marks the first time a major semiconductor company has shipped a CPU explicitly designed for AI agent workloads, potentially establishing a new product category in data center silicon. For agent platform operators and orchestration framework developers, dedicated CPU hardware optimized for tool calls, sandbox execution, and multi-agent concurrency could materially reduce the latency and cost of running autonomous systems at scale. If NVIDIA's performance claims hold under independent benchmarking, Vera could accelerate the deployment density of production agent systems across the cloud ecosystem.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry