智东西（公众号：zhidxcom）
作者 | ZeR0
编辑 | 漠影
智东西11月3日报道，近日，在2026腾讯全球数字生态大会上，英特尔与腾讯云发布基于英特尔至强6+处理器的腾讯云新一代云实例，并全面展示英特尔的芯片产品与云端实践。
全新至强6+处理器拥有最高288颗能效核，单颗处理器最高可支持超过1000个智能体的稳定部署，这也为腾讯云服务器提供更高并发、更低时延、更优能效的云端算力服务。
在至强6+支撑下，腾讯云新一代云实例综合性能较上一代提升30%以上，适配于云原生微服务、高并发Web与智能体任务编排等场景，进一步扩展了腾讯云面向智能体AI工作负载的产品能力。
此外，新一代云实例引入英特尔TDX（信任域扩展），通过硬件级隔离与认证，为智能体构筑“数据可用不可见”的全生命周期可信防线。
面向下一阶段更高密度的任务调度与并发计算，英特尔将推出专为下一代智能体AI基础设施设计的至强处理器（代号Diamond Rapids），实现多任务高负载的吞吐飞跃。
面向智能体AI对基础设施提出的高并发、低时延等需求，英特尔与腾讯云还围绕CPU性能、AI存储方案创新、沙箱拉起和软件生态等话题分享了最新实践，共同探索云端智能体AI基础设施的演进路径。
“面向未来，英特尔期待与腾讯进一步深化‘芯、云、端’一体化协同，共建面向智能体AI的全栈算力底座，让算力各得其所，让智能触手可及，共同推动智能体AI从技术创新走向规模化应用。”英特尔市场营销集团副总裁、中国区总经理郭威说。
▲英特尔市场营销集团副总裁、中国区总经理郭威
一、借助英特尔IAA，对内存解决方案做创新
智能体AI要走向规模化，每一环节都要经得住高密度、高并发的考验。
比如，有限的内存要承载更多任务，KV Cache要在不同层级间高效流动、压缩与调度，沙箱要安全且快速拉起，多模态数据还要被高效处理。
面对这些贯穿计算、存储与执行链路的新要求，英特尔将至强处理器的计算能力、内置硬件加速和软件优化结合起来，与腾讯云共同打通智能体AI运行的关键链路。
▲腾讯全球数字生态大会英特尔展台
智能体AI越深入真实业务，内存就越显珍贵。如何在高密度、高并发的业务场景下，同时满足对密度和时延的双重需求，这是软件方案解决不了的问题，也是腾讯云与英特尔推进内存应用创新的出发点。
据介绍，腾讯团队最初测试了LZ4、ZSTD等软件方案，LZ4在单路解压时，时延尚能接受，多路并发之后，最大时延却曾达到5毫秒，无法满足该业务场景的要求，此外并发增加还带来了内存带宽争抢。
而借助英特尔至强处理器内置的英特尔IAA（内存分析加速器），腾讯云对内存解决方案进行创新，为内存压缩和解压缩引入了硬件加速，满足客户对解压时延极度敏感、必须严格可控的要求。
IAA可以为压缩、解压缩提供硬件加速，具备低时延、低功耗和较低CPU占用等优势。腾讯团队对内核驱动进行了轻量适配，简化相关调用路径，并测试了不同的批量并行配置，最终选择IAA-batch-8。
和软件方案相比，2M页的解压有了IAA加持后，解压时延被拉进微秒级，平均耗时加速16倍、最大耗时加速9倍，在提升内存利用率的同时，实现了更稳定、可预测的响应，让内存不再成为高密度、高并发智能体AI的性能瓶颈。该方案已经过腾讯云验证并已完成业务上线。
二、打造稳定可靠沙箱底座，单机支持350个沙箱并发创建
智能体AI要高效运转，执行环境也要随叫随到。随着智能体AI工作负载开始大规模落地，基础设施面临的另一项重要挑战，是如何在极短时间内批量创建并拉起海量沙箱，高效完成执行环境的编排与隔离。
基于英特尔至强处理器与腾讯云CubeSandbox的深度协同，双方共同打造高并发、极速启动、稳定可靠的沙箱底座。
CubeSandbox是腾讯云开源的弹性智能体沙箱计算平台，承载着智能体沙箱服务、云函数等弹性负载。依托至强6+处理器的计算性能、灵活的扩展能力与可靠的平台能力，CubeSandbox可高效支撑高密度沙箱负载，实现更低的启动时延和更高的资源利用效率。
经腾讯实测，在满足200毫秒SLA要求的前提下，搭载至强6+处理器的CubeSandbox，单机可支持350个沙箱并发创建请求，较其他主流测试平台提升约46%。
面对大量环境同时启动的需求，腾讯云在软件层面做了不少准备。例如，将创建沙箱所需的部分资源提前池化，等到请求到来时直接取用，减少临时创建的开销；对没有修改的内存和磁盘内容尽量共享，降低重复占用。
英特尔与腾讯团队还共同排查内核执行路径，把相关优化合入OpenCloudOS内核，进一步降低并发启动时延。
对于智能体服务，这样的优化有很实际的意义。环境准备得更快，任务也就能更早开始；同一台机器承载更多并发创建请求，能进一步提升资源使用效率；而高负载下保持平稳，则能让这些能力更接近生产环境的要求。
三、充分利用CPU资源，提高多模态任务效率
执行环境要快速就绪，进入真实业务的智能体还要面对另一类挑战：比如快速批量处理具身智能数据等。智能体与现实世界的交互早已超越文本，视频、图像、音频等数据正在大量进入工作流，对数据处理和推理效率提出了更高要求。
针对这一变化，腾讯云多模态计算引擎Xpark应用至强处理器内置AI加速引擎英特尔AMX（高级矩阵扩展），将CPU资源用于小模型推理和轻量级多模态任务，进而提高AI计算吞吐与整体资源利用效率。
Xpark有几项关键能力，包括有非常丰富的高性能多模态算子、异构资源协同加速、英特尔AMX指令集加速、单机算子分布式化实现性能突破等。
在Xpark视频处理工作流中，在iVTAL（英特尔视频转码加速库）的加持下，视频转码性能和开源PyAV相比提升了1.5倍，整体视频处理效率显著提高。
目前，Xpark已正式上线，应用于实际业务场景之中。
围绕智能体不断增长的上下文需求，英特尔还介绍了基于至强处理器的智能体AI的存储平台。
面向大容量、高性能、高能效、易部署的存储需求，英特尔通过高速I/O，QAT、DSA等内置加速器、UINTR+TPAUSE专用指令等软硬件协同优化，解决KV Cache的流转和复用效率等问题，让存储从放得下走向流得动和效率高。
四、打造系统级产品组合，加速智能体规模化落地
英特尔为客户打造了覆盖CPU、GPU、存储和网络的系统级方案，产品组合涵盖至强、以太网控制器和GPU D系列等。
其中，英特尔GPU D系列（代号Crescent Island）专为持续运行的内存密集型AI推理工作负载而打造：基于Xe3P架构，在350瓦功耗范围内，可为用户提供高达480GB的LPDDR5X内存、1.5TB/s内存带宽以及1.3PFLOPS FP8算力，让大容量推理更易部署。
要让硬件性能真正转化为开发者的生产力，开放、易用的软件生态同样关键。英特尔坚持AI软件栈开源并优先贡献至上游社区，覆盖PyTorch、vLLM、SGLang等主流框架，以及Triton、oneDNN、oneCCL等计算与通信库，帮助主流模型实现开箱即用。
此外，针对已有代码的跨平台迁移，英特尔引入基于大语言模型的AI Agent，结合专业Skill能力，在理解算法意图的基础上原生重构代码，并完成编译、验证与性能调优的全流程闭环，让跨架构迁移从“能跑”走向“跑好”。
同时，英特尔将算力与智能体能力从云端延伸至终端、边缘。搭载WorkBuddy、Marvis的AI PC和基于英特尔酷睿Ultra处理器的车载AI Box等端侧应用集中亮相，生动呈现了端云协同的端侧智能体验，让智能体AI走进千家万户。
▲腾讯全球数字生态大会英特尔展台
结语：英特尔以至强为算力中枢，构建智能体AI算力底座
英特尔市场营销集团副总裁、中国区总经理郭威谈道，从数据中心到智能终端，英特尔与腾讯持续拓展合作边界，以至强6/6+夯实云端算力底座，以智能体PC拓展端侧AI应用空间，通过模型与Agent深度协同优化释放全栈智能潜力，并依托开放生态汇聚更广泛的创新力量，共同贯通智能体AI从基础设施、模型到应用的关键链路。
从软件测试遇到瓶颈，到引入硬件加速，再到驱动适配和业务上线，英特尔与腾讯云合作的案例展示了一条完整的优化路径。面对智能体从技术探索走向全场景生产力的趋势，依托覆盖算力、存力、软件优化等在内的系统级创新，英特尔正以至强为算力中枢，以开放、包容的全栈平台能力，构建智能体AI时代的高性能智算底座。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
Author ZeR0
Edit |
Intel and Tencent cloudZ publish are examples of a new generation of Tencent clouds based on Intel to Strong 6+ processors, and a full display of Intel's chip products and cloud practice.
New to strong 6+ processors have up to 288 energy-efficient cores and a single processor can support the stable deployment of more than 1,000 AI agents, which also provides higher co-production, lower latency, more energy-efficient cloud computing services for Tencent cloud servers.
Supported by a maximum of 6+, the Tencent cloud cloud new-generation cloud complex can increase by more than 30 per cent over the previous generation. It is suitable for micro-services in clouds, high composing Web and AI agent missions, and further expands the product capacity of the Tencent cloud for AI agent AI.
In addition, a new generation of cloud examples was introduced into IntelTDX (trust domain extension) to construct a credible life-cycle defence for AI agent through hardware-level isolation and authentication.
For the next stage of mission deployment and simultaneous calculation with higher density, Intel will launch a powerful processor (named Diamond Rapids), designed for the next generation of AI agent AI infrastructure, to achieve a multi-tasking high-load perv.
Intel and Tencent clouds have also shared recent practice on topics such as CPU performance, AI storage programme innovation, sandbox lifting and software ecology, exploring the evolution of cloud-end AI agent AI infrastructure, addressing the high and low AI agent AI needs for infrastructure.
“For the future, Intel looks forward to further deepening the `core, cloud, end' integration synergy with Tencent, which will create an all-in-the-mill base for AI agent AI, so that it will be able to take advantage of it, so that intelligence can reach it, and so that AI agent AI can move from technological innovation to scale applications.” Guo Wai, Vice-President of Intel Marketing Group and General Manager of China District, said:
Kwok Wei, Vice-President of the Intel Marketing Group and General Manager of China District
I. Innovative memory solutions through IntelIA
AI agent AI has to go to scale, and each link has to stand up to a high-density, high-intensity test.
For example, limited memory is required to carry more tasks, KV Cache is required to move, compress and move efficiently between different levels, sandboxes are required to be safely and rapidly pulled up and multimodular data processed efficiently.
In the face of these new requirements, which cut across the computation, storage and implementation links, Intel combines the computational capability of the powerful processors, the acceleration of built-in hardware and software optimization, with the Tencent cloud to connect to the AI agent AI operations.
Intel exhibition of the Tencent Global Digital Ecology Congress Station
The more AI agent AI goes into real business, the more valuable the memory is. The question of how to meet the dual demand for density and latency in a high-density, high-symmetric business landscape is a problem that cannot be solved by the software programme and is the starting point for Tencent clouds and Intel to advance memory applications.
It was described that the Tencent team had initially tested software programs such as LZ4, ZSTD, and that when LZ4 was decompressed on a single-way basis, latency had been able to accept it and that, after the multiple circuits, the largest latency had reached 5 milliseconds before it could not meet the requirements of the business scene, and that the increase had also resulted in a RAM bandwidth contest.
With the IntelIA (RAM Analytic Accelerator) embedded in Intel to Power Processors, the Tencent cloud innovates memory solutions, introducing hardware acceleration for memory compression and decompression, and meeting the client ' s extreme sensitivity to depression latency, which must be strictly controllable.
IAAs can provide hardware acceleration for compression and decompression, with advantages such as low latency, low power consumption and lower CPU occupancy. The Tencent team adapted the kernel drive lightly, simplified the associated call paths and tested different batch parallel configurations, eventually choosing IAA-batch-8.
Compared with the software program, 2M-page decompression with IAA, the decompression of latency was pulled into microseconds, with average latency accelerating 16 times and the maximum time-consuming 9 times, while increasing the utilization of memory, achieving a more stable and predictable response, so that memory is no longer a high-density, high-cotermine performance bottleneck. The programme has been validated by Tencent clouds and has been operationalized.
II. Building stable and reliable sandbox floors with a single machine supporting 350 sandboxes and creating them
AI agent AI is to function efficiently and the implementation environment is on call. As AI agent AI loads begin to land on a large scale, another important challenge for infrastructure is how to create and lift large-scale sandboxes in very short periods of time and efficiently organize and isolate the implementation environment.
Based on the deep synergy between the Intel to Power Processor and the Tencent cloud CubeSandbox, the two sides have worked together to create high-speed, high-speed, stable sandbox bottoms.
CubeSandbox is the elastic AI agent sandbox calculation platform for the open source of Tencent clouds, carrying flexible loads such as AI agent sandbox services, cloud functions, etc. CubeSandbox can efficiently support high-density sandbox loads, achieve lower start-up latency and higher resource efficiency, depending on the computational performance of a strong 6+ processor, flexible extension capacity and reliable platform capacity.
Based on Tencent, and subject to 200 msSLA requirements, CubeSandbox, which is loaded to a 6+ processor, can support 350 sandboxes and issue requests, an increase of approximately 46 per cent over other mainstream testing platforms.
Tencent clouds were prepared at the software level in the face of the need for simultaneous start-up in a large number of environments. For example, some of the resources required for the creation of sandboxes will be pooled in advance and will be taken directly upon request to reduce the costs of temporary creation; memory and disk content without modifications will be shared as much as possible and duplicate occupancy will be reduced.
Intel, together with the Tencent team, is also conducting a mapping of the internal nuclear implementation path, integrating the optimization into the OpenClaudOS kernel, further lowering and launching latency.
For AI agent services, such optimization has practical implications. The faster-prepared environment and the earlier start of the mission; the greater availability of the same machine and the increased number of requests for creation could further increase resource efficiency; and the maintenance of stability under high loads could bring these capabilities closer to the requirements of the production environment.
III. Making full use of CPU resources to increase the efficiency of multi-model missions
The implementation environment needs to be ready quickly, and AI agent, which enters the real business, faces another type of challenge: fast bulk processing of smart data. The interaction between AI agent and the real world has long gone beyond text, and video, image, audio and other data are entering workflow in large quantities, placing greater demands on data processing and reasoning efficiency.
In response to this change, the Tencent cloud multi-module computing engine, Xpak, was applied to the strong processor to incorporate the AI Accelerator engine, Intel AMX (Advanced Matrix Extension), which used CPU resources for small model reasoning and lightweight multi-modular tasks, thus increasing the efficiency of AI computing throughput and overall resource utilization.
Xpark has several key capabilities, including a very rich high-performance multimodular calculator, syndromic acceleration of isomer resources, acceleration of the Intel AMX command set, and the achievement of performance breakthroughs in mono-calculator distribution.
In Xpak video processing workflow, with the addition of iVTAL (Intel video transcode acceleration library), video trans-code performance and open source PyAV increased 1.5 times and overall video processing efficiency improved significantly.
At present, Xpak is officially online and is being used in the operational landscape.
Intel also presented a storage platform based on AI agent AI, a powerful processor, around the growing demand for AI agent.
For high-capacity, high-performance, energy-efficient, deployable storage needs, Intel is able to move storage from lower to more efficient by synergizing software and hardware, such as high-speed I/O, QAT, DSA, and UINTR+TPAUE commands.
Building a system-level product mix to accelerate the size of AI agent Land
Intel has developed system-level programmes for clients covering CPU, GPU, storage and network, with product combinations ranging from strong to Ethernet controllers to GPU D series.
Among them, the Intel GPU D series, codenamed Creative Island, was designed for the ongoing memory-intensive AI reasoning load: based on the Xe3P structure, with up to 480GB LPDDR5X memory, 1.5TB/s memory bandwidth and 1.3 PFLOPS FP8 computing power, large capacity reasoning is available to users within 350 WW.
Open and easy-to-use software ecology is also crucial to making hardware performance truly productive for developers. Intel adheres to the open source of AI software and prioritizes its contribution to upstream communities, covering mainstream frameworks such as PyTorch, v LLM and SSGLang, as well as the Triton, oneDNN, oneCCL, etc., computing and communication libraries to help mainstream models get ready to open.
In addition, in response to cross-platform migration of existing codes, Intel introduced AI Agent, based on a large-language model, combining professional Skill capabilities, with original re-constructing codes based on an understanding of algorithm intent, and completed the compilation, validation and performance-modified full-process closed loops to allow cross-structure migration from “capable” to “good”.
At the same time, Intel extends the algorithm and AI agent capability from cloud end to terminal, edge. The side applications of the Al PC and the AI Box, based on the Ultra processor based on Intel ' s idiosyncrasy, brought together a live display of the AI agent test on the side of the cloud, allowing AI agent AI to enter thousands of homes.
Intel exhibition of the Tencent Global Digital Ecology Congress Station
Concluding remarks: Intel is the center of power, building AI agent AI
Guo Wei, Vice-President of the Intel Marketing Group and Managing Director of the China District, spoke of the ongoing expansion of the cooperation boundary from data centres to smart terminals, and of Intel and Tencent to a strong 6/6+ cloud-based computational base to expand the side AI application space with AI agentPC, optimize the release of the full range of intellectual potential through modeling in collaboration with Agent depth, and build on open ecology to bring together broader innovation forces to connect AI agent AI from infrastructure, models to applications.
From bottlenecks in software testing to the introduction of hardware acceleration to driving adaptation and operational go-live, the case of Intel working with Tencent clouds shows a complete optimisation path. In the face of the trend of AI agent from technological exploration to the productivity of the whole scene, based on system-level innovation, including coverage of computing power, viability, software optimization, Intel is using it as a powerful center of computation, building the high-performance intellectual base of the AI agent AI era with an open and inclusive whole-house platform capacity.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.