
AI 已成为不可阻挡的巨轮,推动着科技行业内外的技术战略。随着全球服务器市场在单一年份内增长 30%,于 2025 年第一季度达到 1226 亿美元,IDC 现在认为将 AI 驱动的投资重新归类为“从周期性到持久性”是稳妥的。这家分析公司预计,到明年,推理将占计算需求的 70% 以上,并补充称其在边缘端的增长最快。
AI has become a juggernaut, driving technical strategy in the tech sector and outside of it. With the worldwide server market growing 30 percent in a single year to hit $122.6bn in the first quarter of 2025, IDC now feels safe recategorizing AI-driven investment "from cyclical to durable". The analyst company expects inference to account for over 70 percent of computing demand by next year, adding that it's growing fastest at the edge.
为何企业纷纷自建基础设施所有这些对 AI 基础设施的投资,一度可能主要来自超大规模云服务商,但这种情况正在迅速改变。IDC 补充道,采用正在迅速蔓延至公共云服务提供商之外的企业环境。企业渴望在内部运行智能体 AI。公司拥有自己的 AI 基础设施而非将其外包给云端,有许多理由。其中之一是数据控制权,特别是在金融、医疗、制造和公共部门等有主权数据要求的受监管行业。另一个原因在于,企业需要通过可调优的内部系统交付 Token,从而获得可预测的成本。
Why enterprises are building their own All this investment in AI infrastructure might once have come mainly for hyperscalers, but that's changing quickly. Adoption is spreading quickly beyond public cloud service providers into enterprise environments, IDC adds. They're eager to run agentic AI in-house. There are plenty of reasons for companies to own their own AI infrastructure rather than offloading it to the cloud. One of these is data control, especially in regulated sectors such as finance, healthcare, manufacturing, and the public sector that have sovereign data requirements. Another revolves around the need for predictable costing from tokens delivered on tunable in-house systems.
那么,公司在选择供应商时应关注什么?对于那些面临巨大压力、必须为业务交付 AI 计划的公司而言,尤其是在内存和 CPU 供应受限的时期,交付速度将成为强大的驱动力。生命周期服务也将名列前茅,适用于那些需要帮助管理从电力到推理大语言模型(LLM),再到使用这些模型的工作负载等一切事务的公司。这种对专业知识的需求不太会停止;随着企业逐渐掌握这项仍在不断演进的技术,其 AI 战略将发生变化。
So what should companies look for in a supplier? Delivery speed will be a powerful driver for many who are under the gun to deliver AI initiatives for the business, especially during periods of constrained memory and CPU supply. Life cycle services will also come near the top of the list, for companies that need help managing everything from the power through to the reasoning LLM models and the workloads that use them. That need for expertise is unlikely to stop; companies will undergo changes in AI strategy as they get to grips with a still-evolving technology.
从桌面到数据中心ASUS 预计,这种 AI 激增将在数据中心持续,但也会蔓延至其他领域,特别是边缘位置(尤其是工业用例)和开发者的桌面,因为将为企业赋能的应用程序正在那里诞生。
From the desk to the datacenter ASUS expects this AI surge to continue in the datacenter, but elsewhere too, specifically in edge locations (especially industrial use cases) and at developers' desks, as the applications that will power the enterprise get made.
它希望成为人们在 AI 领域各个环节中,硬件和服务的首选。在 10 月 21 日至 22 日于柏林举行的 NVIDIA GTC 大会上,它发布了一项名为“企业智能体 AI 解决方案”的全栈 AI 产品。这是一个三层架构,包含:ASUS AI 服务器与基础设施:其 AI 服务器产品线,包括多款新机型。ASUS AI 软件栈:用于集成和控制 AI 工作流的软件。
It wants to be top of peoples' list for hardware and services at all these places along the AI continuum. At NVIDIA's GTC show in Berlin on Oct 21-22, it announced a full-stack AI offering that it's calling its Enterprise Agentic AI Solution. It's three-layer affair, featuring: ASUS AI Server and Infrastructure: Its selection of AI servers, including several new units. ASUS AI Software Stack: Its software for integrating and controlling AI workflows.
ASUS AI HUB:将所有内容串联起来的服务运营体系。这些将华硕原有的面向企业 AI 用户的资产与一些新资产整合在一起。华硕高管表示,鉴于市场正在发生变化,现在是推出搭载新硬件的端到端产品的合适时机。“华硕服务器帮助企业构建高性能智能体 AI 基础设施,能够同时支持复杂任务执行和大规模 AI 智能体部署,加速企业采用智能体 AI,”华硕资深副总裁兼基础设施解决方案事业群总经理 Paul Ju 表示。
ASUS AI HUB: A services operation that ties it all together. These bring together some existing assets that ASUS had for enterprise AI users, along with some new ones. The time is right for an end-to-end offering with new hardware because of the way the market is shifting, say ASUS execs. "ASUS servers help enterprises build high-performance agentic AI infrastructure capable of supporting both complex task execution and large-scale AI agent deployment, accelerating enterprise adoption of agentic AI," says Paul Ju, senior vice president and head of infrastructure solution BG at ASUS.
生成式 AI 最初引发了需求激增,而智能体 AI 的出现则让算力需求进入“超驱动”状态。Ju 解释说,关注点正从 GPU 扩展开来。“历史上,AI 服务器主要围绕 GPU 算力进行优化。但随着企业 AI 向智能体 AI 转移,工作负载越来越多地围绕一个持续的‘智能体循环’展开,这高度依赖 CPU,”他说。该循环包含任务规划、代码执行、外部工具调用,以及反思或纠错工作。
While generative AI initially spiked demand, the advent of agentic AI has sent compute demand into overdrive. Ju explains that the focus is expanding from the GPU. "Historically, AI servers have been optimized primarily around GPU compute. But as enterprise AI shifts toward agentic AI, workloads increasingly revolve around a continuous ‘agent loop’ that is heavily CPU-dependent," he says. That loop takes in task planning, code execution, external tool invocation, and reflection or correction work.
他补充说,CPU 在执行工作中的角色增强,正在传统 AI 服务器中造成瓶颈。第一个瓶颈是单线程性能,这导致 GPU 在等待 CPU 完成任务时处于空闲状态。第二个瓶颈在于内存带宽。当按核心管理内存带宽时,大量智能体同时运行会拖慢速度。
The CPU's increased role in this execution work is creating bottlenecks in conventional AI servers, he adds. The first such sticking point is single-thread performance, which leaves GPUs idling while they wait for CPUs to complete. The second lies in memory bandwidth. When this is managed on a per-core basis, it can slow things down when many agents run at once.
许多 CPU 的多芯片设计在 AI 环境中也会带来劣势,因为跨芯片访问数据会引入延迟,Ju 认为。核心与缓存位于不同芯片的 CPU,在数据必须在它们之间传输时会遭受延迟。维护缓存一致性会加剧这一问题。最后,还存在安全风险:他警告称,部分 CPU 缺乏硬件级沙箱隔离,使得提示词注入有可能触达核心系统。客户在应对不同的 AI 应用时也带来了自身的挑战。
The multi-die design of many CPUs also penalizes them in AI environments, because accessing data across dies introduces latency, Ju argues. CPUs with cores and cache on different dies suffer latency when data has to cross between them. Maintaining cache coherency exacerbates that problem. Finally, there's a security risk: a lack of hardware level sandbox isolation on some CPUs make it possible for prompt injection to reach core systems, he warns. Customers also bring their own challenges as they grapple with different applications for AI.
Ju 补充道,评估 AI 服务器选项的企业必须将其服务器平台与预期的 AI 工作负载概况相匹配。训练一个用于批量分类保险理赔的监督学习模型,所需的技术资源与运行一个低延迟推理员工聊天机器人截然不同。华硕在柏林展出的产品华硕希望帮助客户解决此类问题,无论它们出现在何处。基于这一考量,公司在 GTC 上为观众准备了丰富的展品。
Enterprises evaluating AI server options have to match their server platforms to the projected AI workload profiles, Ju adds. Training a supervised learning model to batch-classify insurance claims takes different technical resources to running a low-latency reasoning employee chatbot. What ASUS took to Berlin ASUS wants to help customers solve problems like these wherever they come up. With this in mind, the company had plenty of goodies for onlookers at GTC.
从数据中心层面来看,公司发布了基于 NVIDIA HGX Rubin NVL8 八 GPU 服务器主板的单机系统和全机柜系统。其中包括一款 3U 混合冷却系统 XA NR11-E12L,以及一款 2U 100% 液冷系统 XA NR11-E12,两者均面向大规模 AI 推理工作负载。这两款 Rubin 系统针对不同的使用场景。3U 混合液冷系统在 8 GPU 基板上采用芯片级直接液冷,同时在滑架上搭载风冷双路至强 6 处理器。
Starting at the datacenter, the company announced both single and full-rack systems built around the NVIDIA HGX Rubin NVL8 eight-GPU server board. It has a 3U hybrid-cooled system, the XA NR11-E12L, along with the 2U 100 percent liquid-cooled XA NR11-E12, both targeting large AI reasoning workloads. The two Rubin units target different use cases. The 3U hybrid liquid cooling system uses direct-to-chip liquid cooling on the 8GPU baseboard, along with air-cooled dual Xeon 6 chips on a sled.
公司将其定位为迈向 AI 工厂硬件的平滑过渡路径。2U 机型凭借 100% 液冷设计,专为最大化 PUE 和算力密度而打造。它还采用无线缆板对板设计,以保持占地面积紧凑。该套件性能也将强劲,NVIDIA HGX Rubin 每 GPU 提供 800Gbit 的集成带宽。华硕通过采用免工具 CPU 和 GPU 滑架,宣传这些机型的易维护性。
The company pegs this as the easy transition path into AI factory hardware. The 2U unit, with its 100 percent liquid cooling, is tailored for maximum PUE and compute density. It also uses a cable-free board-to-board design to keep its footprint low. The kit will be speedy, too, with NVIDIA HGX Rubin offering 800Gbit of integrated bandwidth per GPU. ASUS is promoting the serviceability of these units through their use of tool list CPU and GPU sleds.
该公司声称,针对这些设备的维护操作可节省高达 50% 的时间,意味着每次服务的停机时间缩短,最终降低总拥有成本。为 NVIDIA HGX Rubin 快速交付ASUS 将其硬件主张构建在快速交付之上,以便客户能迅速让硬件投入运行,支持其 AI 计划。这种速度源于交付前的验证工作,并借助预验证的机架参考架构实现。
It claims up to a 50 percent time saving on maintenance operations for these units, meaning less downtime per service and ultimately a lower total cost of ownership. Delivered fast for NVIDIA HGX Rubin ASUS is structuring its hardware proposition around speedy delivery so that customers can get hardware up and running to support their AI initiatives. That speed comes from validation carried out before delivery with the help of a pre-validated rack reference architecture.
NVIDIA MGX 服务器助力灵活构建秉持针对不同 AI 工作负载提供多种设计的理念,ASUS 还支持 NVIDIA 模块化 MGX Blueprint 服务器构建方案。其卖点在于灵活性,可混合搭配 GPU、CPU 和网络选项,适用于企业推理、智能体及视觉计算。这是一种模块化路径,适用于不需要 8-GPU HGX 基板算力的工作负载,覆盖大量企业推理和智能体用例。
NVIDIA MGX Server for flexible builds In keeping with its focus on multiple designs for the right AI workload, ASUS also supports NVIDIA's modular MGX Blueprint for server builds. The selling point here is flexibility with a mixture of GPUs, CPUs, and networking options for enterprise inference, agents, and visual computing. This is the modular route for workloads that don't need the power of an 8-GPU HGX baseboard, and it covers a large gamut of enterprise inference and agentic use cases.
ASUS 展示了 MGX 服务器,包括配备八张液冷 NVIDIA RTX PRO 6000 Blackwell Server Edition GPU 的 6U XA P8A-E14AXL。其 2U 同胞机型 XA P4N-E2 搭载 NVIDIA Vera 处理器,面向智能体工作负载。面向边缘与开发者的解决方案ASUS 展台还展示了 ExpertCenter Pro ET900N G3,搭载配备 748GB 统一内存的 GB300 Grace Blackwell Superchip 开发者桌面,以及基于 NVIDIA DGX Spark、面向智能体工作的 Ascent GX10。
ASUS showcased MGX servers including the 6U XA P8A-E14AXL with its eight liquid-cooled NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. A 2U sibling, the XA P4N-E2, sports NVIDIA Vera processors for agentic work. Solutions for the edge and developers The ASUS booth also featured the ExpertCenter Pro ET900N G3, powered by the GB300 Grace Blackwell Superchip developer desktop with 748GB of coherent memory, and the Ascent GX10 based on NVIDIA DGX Spark for agentic work.
现场还展出了 ProArt GR1X 迷你主机,以及 ProArt P16 和 P14 笔记本——首批搭载 NVIDIA RTX Spark 的笔记本,面向开发者和创作者。针对边缘部署,展示了 PE3000N,这是一款基于 NVIDIA Jetson T5000 的坚固型边缘 AI 计算机,适用于工业环境等用例。该产品发布于一年前,可提供高达 2,070 FP4 TFLOPS 的峰值 AI 性能,配备 128GB 内存。
Also on show were the ProArt GR1X Mini PC and ProArt P16 and P14 laptop - the first laptops powered by NVIDIA RTX Spark - aimed at both developers and creators. For edge installations, it showed the PE3000N, its rugged NVIDIA Jetson T5000-based edge-AI computer aimed at use cases including industrial environments. Announced a year ago, this delivers up to 2,070 FP4 TFLOPS of peak AI performance with 128GB of memory.
面向工业 AI 环境的则是 PE6000G,一款工作站级 AI 与仿真平台。在数字孪生中演练 AI 工厂只有有电力驱动,设备才能支撑 AI 工厂。华硕一直在与英伟达合作,采用其 DSX AI Factory 平台及数字孪生系统,用于 AI 工厂环境中的硬件与电力布局。DSX 支持搭载英伟达 Vera Rubin 处理器系列的设备,该系列支持干式冷却器,并可在 45 摄氏度温水进水条件下实现更低的 PUE(电源使用效率)。
Targeted at industrial AI environments was the PE6000G, a workstation-grade AI and simulation platform. Rehearsing the AI factory in a digital twin Equipment only supports an AI factory when there is power to run it. ASUS has been taking steps with NVIDIA as a partner here, adopting its DSX AI Factory platform with digital twin system for laying out hardware and power in AI factory environments. DSX supports equipment using the NVIDIA Vera Rubin processor line, which supports dry coolers and a lower PUE with a 45 degree centigrade warm water inlet.
华硕配置的机柜在 DSX 中成为可仿真的资产,采用其 OpenUSD 共享文件格式。这涵盖了工程数据以及用于测量实际功耗和冷却需求的几何数据。华硕希望将冷却作为 AI 工厂原始设计计划的一部分,而非设施建设的事后补充。因此,DSX 合作使实施团队能够在数字孪生模型中对水温、流量、压力、CDU(冷却分配单元)布局和管道进行建模。
ASUS configured racks become simulation-ready assets inside DSX, adopting its OpenUSD shared file format. That covers engineering data alongside the geometry for measuring real power draw and cooling requirements. ASUS wanted to target cooling as part of the original design plan for the AI Factory, not as a facility afterthought. So, the DSX partnership allows implementation teams to model water temperature, flow, pressure, CDU placement, and piping within the digital twin model.
他们利用华硕基础设施部署中心,转换来自规划资源的信息,例如英伟达 Omniverse DSX Blueprint 框架。构建更快,运维更易数字孪生系统不仅用于设计和部署,还旨在上线后持续发挥作用,以便 AI 工厂运营商在接触硬件前先在仿真中验证变更。华硕控制中心监控基础设施,而 DSX 交换链路连接 IT 与运营技术数据。
They translate information from the planning resources that go around this, such as the NVIDIA Omniverse DSX Blueprint framework, using the ASUS Infrastructure Deployment Center. Faster to build, easier to run The digital twin system isn't just for design and setup. It's supposed to keep working after go-live so that AI factory operators can simulate changes before touching hardware. The ASUS Control Center monitors the infrastructure while DSX exchange links IT and operational technology data.
运营商可在仿真中验证变更,随后再切换上线。华硕此举的目标清晰:协助企业客户应对内部 AI 机遇,同时认识到其工作负载的性能特征和物理部署位置将各不相同。随着企业不断深化对 AI 及其在组织中定位的理解,它们将需要在从数据中心到边缘、再到开发者工位的基础设施规划与培育上保持敏捷。
Operators can validate changes in the simulation before flipping the switch to go live. ASUS' goal here is clear: to assist enterprise customers as they tackle in-house AI opportunities while acknowledging that the performance profile and physical locations of their workloads will vary. As companies continue to evolve their understanding of AI and where it fits within the organization, they will need to be agile in how they plan and nurture their infrastructure from the datacenter out to the edge, and to the developer's desk.
柏林的发布将这些内容整合成一个整齐的方案包,为那些在智能体化愿景方面需要帮助的组织提供支持。了解更多关于 ASUS NVIDIA HGX Rubin NVL8 系统的信息:XR R1I12B-M1BLFS | ASUS 服务器 由 ASUS 赞助
The Berlin announcements tie it together in a neat package for organizations that need help with their agentic aspirations. Learn More about ASUS NVIDIA HGX Rubin NVL8 Systems: XR R1I12B-M1BLFS | ASUS Servers Sponsored by ASUS