您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：腾讯云实测英特尔至强6+平台，内存解压平均时延改善16倍，单机支持350个沙箱并发创建请求。
作者 | 陈悦琳
编辑 | 包永刚
一个智能体能顺利完成任务，不代表成百上千个智能体就能同时高效运行。
对于承载大量智能体任务的云服务器而言，除了提升计算速度，如何在任务集中到来时迅速准备好执行环境、怎样减少暂时闲置的任务对内存资源的占用，也成为亟待解决的问题。
对此，在一年一度的腾讯全球数字生态大会上，英特尔与腾讯云联合发布了基于至强6+处理器的新一代云实例，围绕任务启动效率、内存资源利用率等关键因素进行优化后，其综合性能较上一代提升30%以上。
从此前搭载至强6处理器的S9、S9e、S9pro，到如今的新一代云服务器（CVM），英特尔与腾讯云的合作并未停留在处理器性能的升级上。面对当前模型调用量持续攀升、智能体开始在实际业务中承担更多工作等趋势，CPU如何与操作系统、内存等环节协同，正成为云服务器的新考题——而这也是双方此次联合优化的重点。
288核至强6+处理器，支撑Agent部署还需跨过内存关
AI应用正攀升至全新的量级。中国信通院发布的《大模型词元服务性能监测报告（2026年）》显示，2026年6月，国内日均Token消耗量已达到约500万亿，是去年12月的近4倍。
这场有关AI使用的“狂欢”不止发生在中国，还在以Agent这一应用形态席卷全球。在麦肯锡覆盖全球97个国家的调查中，在年营收超过10亿美元的企业受访者中，40%表示其所在企业已在至少一个业务职能中规模化部署Agent，高于2025年的27%。
与传统的一问一答不同，智能体能够将一项任务拆解为多个步骤，反复调用模型、使用工具、执行代码，并根据反馈调整行动。这不仅可能带来更多Token消耗，也让CPU需要承担更多模型推理之外的工作。
据英特尔市场营销集团副总裁、中国区总经理郭威观察，今年上半年行业仍在以1:8、1:4等比例讨论CPU与GPU的配置关系；进入下半年，业界对CPU需求继续增长的预期愈发明显，甚至有人提出未来可能向1:1演进。
在服务器CPU产品布局上，英特尔早在2023年就明确了性能核（P-core）与能效核（E-core）双线并行的方向，以满足不同工作负载对计算性能、核心密度和能效的需求。其中，性能核侧重计算性能，适合对单核性能要求较高的工作负载；能效核则强调核心密度和能效，更适合大规模并行、高吞吐量的计算任务。
今年6月，英特尔推出采用Intel 18A制程的至强6+处理器（代号Clearwater Forest）。该处理器最高集成288个能效核，支持12条内存通道，DDR5最高传输速率达到8000MT/s。英特尔还披露，单颗至强6+处理器最高可支持超过1000个智能体稳定运行。
不过，CPU核心数量增加、内存传输速度提高，并不意味着服务器就能同比例承载更多智能体。英特尔技术专家在分析智能体运行负载时发现，CPU使用呈现明显的突发性——智能体在启动执行环境、执行任务时，会产生短时的计算需求高峰，但从单个Agent的运行过程来看，CPU整体利用率总体并不高。
与之形成对比的是内存。英特尔专家同时指出，即使智能体暂时没有计算任务，其运行状态仍可能持续占用内存。其在现场算了一笔账：假设每个Agent分配2GB内存，一份vCPU资源运行4个Agent，就需要8GB内存；而在一颗256核CPU、12条内存通道、每通道配置128GB内存的示例中，平均每个vCPU对应的物理内存容量只有6GB。
这意味着，当云厂商希望通过资源复用提高智能体部署密度时，内存容量可能比CPU计算能力更早成为限制因素。因此英特尔与腾讯云还将此次基于至强6+的合作进一步延伸至操作系统内核、资源调度和内存管理等环节，尝试将硬件能力转化为实际业务中的性能提升。
内存要省下来，沙箱也要起得快
TrendForce集邦咨询数据显示，2026年第二季度服务器DRAM平均合约价环比上涨53%—58%。尽管涨势有所放缓，该机构仍预计第三季度价格将继续上涨13%—18%。
面对持续增加的采购成本，对于希望在一台服务器上运行更多Agent任务的云厂商而言，想办法让现有内存容纳更多数据迫在眉睫。
将暂时不需要频繁访问的内存数据压缩存放、等到需要时再解压使用，是提高服务器内存利用率的关键路径之一。但这带来的新问题是，在节省空间的同时，压缩和解压同样需要消耗计算与内存访问资源，尤其在大量任务同时请求解压缩时，额外的时延可能拖慢业务响应。腾讯云TencentOS Server产品研发团队就曾尝试利用LZ4软件方案进行内存压缩，但发现多路并发解压缩的最大时延长达5毫秒，无法满足实际业务需求。
为解决这一问题，腾讯云转向英特尔至强CPU内置的IAA（内存分析加速器），通过专用硬件执行压缩和解压缩任务，减少对CPU计算资源的占用，并改善处理时延。在此基础上，腾讯云团队还对IAA驱动进行了适配，简化不必要的调用路径，并测试了多种批量并行配置，最终选择IAA-batch-8方案。
测试结果显示，IAA的优势不仅在于缩短解压缩时间，还体现在高并发下的稳定性。当并发规模从32路增加至256路，LZ4软件方案的解压缩时延随之上升，而IAA-batch-8的时延基本保持稳定。在更接近实际业务的综合压力测试中，针对2MB内存大页的解压缩，相较软件方案，IAA的平均时延和最大时延分别改善约16倍和9倍，且都处于微秒级别。
除了内存资源外，沙箱创建效率也是影响智能体高密度部署、高并发执行的重要因素。
所谓沙箱（Sandbox），就是为智能体执行代码、调用工具提供的隔离环境，避免不同任务相互干扰。腾讯云CubeSandbox团队指出，在部分大模型强化学习场景中，一轮训练可能需要启动成千上万甚至数十万个沙箱，用于模型评测和推理。如何在短时间内完成大量执行环境的创建，同时控制资源占用，就成了云厂商需要解决的问题。
然而，不同的隔离技术各有局限：传统容器启动较快，但通常共享服务器的操作系统内核；完整虚拟机拥有独立内核，却需要占用更多资源。因此，腾讯云CubeSandbox采用微型虚拟机（MicroVM）、快照和资源复用等技术，使多个沙箱能够共享相同的基础资源，仅对发生变化的部分进行独立存储，从而减少重复的内存占用。
在此基础上，英特尔工程师还针对沙箱创建过程中涉及的部分操作系统内核执行路径进行优化，并将相关改进引入腾讯等企业共同发起的开源操作系统OpenCloudOS内核，进一步降低大量沙箱并发启动时的时延。
现场披露的测试数据显示，在满足200毫秒服务等级目标（SLA）的条件下，搭载至强6+处理器的服务器在运行CubeSandbox时，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%。
更值得关注的是负载叠加后的表现。腾讯云实测发现，即便服务器同时承载其他任务、整体负载上升，至强6+仍保持相对稳定的沙箱并发创建能力。
不过，英特尔并不认为现有产品已经能够完全满足智能体的运行需求。英特尔技术专家在现场表示，智能体应用与软硬件生态之间仍有进一步优化的空间，团队基本每周都会将与合作伙伴开展实际负载优化时积累的经验反馈给后端CPU设计团队，为后续产品的改进提供参考。
双方的合作也将延续至下一代产品。据腾讯云高级产品经理夏士悦透露，腾讯云计划继续与英特尔合作开发基于下一代至强处理器Diamond Rapids的云服务器实例。按照英特尔此前公布的产品规划，Diamond Rapids将采用Intel 18A-P制程，最高集成256个CPU核心，支持16通道内存和128条PCIe 6.0通道，并引入高级性能扩展（APX）及增强型高级矩阵扩展（AMX）等技术。
雷峰网雷峰网(公众号：雷峰网)
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
为了您的账户安全，请验证邮箱
请验证您的邮箱
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。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: Tencent cloud measurements to the Intel to 6+ platform, average memory decompression latency improved 16 times, single machine supported 350 sandboxes and sent requests for creation.
I'm the author.
Edit | Pao Yonggang
A AI agent can successfully carry out its mission, which does not mean that hundreds of Z42Zs can operate efficiently simultaneously.
For cloud servers carrying a large number of AI agent missions, in addition to increasing the speed of calculation, the question of how to rapidly prepare for the implementation of the environment when the tasks come together and how to reduce the use of memory resources for temporarily idle tasks is also an urgent one.
In response, at the annual Tencent Global Digital Ecology Congress, Intel and Tencent Clouds combined publish, a new generation of cloud-based examples based on up to the strength of 6+ processors, optimized around key factors such as mission start-up efficiency and the utilization of memory resources, resulting in more than 30 per cent more integrated than the previous generation.
From S9, S9e, S9pro, which used to carry a 6-manufacturing processor, to the new generation of cloud servers (CVM), Intel ' s cooperation with Tencent cloud did not stop at upgrading the processor performance. In the face of trends such as the continuing growth of the current model calls and the beginning of AI agent to undertake more work in the field, how CPU is working with the operating systems, memory, etc., is becoming a new issue for cloud servers — which is the focus of this joint optimization.
Zoom to 6+ processor 288, which supports Agent's deployment, will need to cross memory Off
The AI application is climbing to a completely new scale. In June 2026, the national average daily consumption of Token reached about 500 trillion, almost four times the rate last December.
This "celebrity" about AI's use is not only happening in China, but is also sweeping the globe in the form of the Agent application. In a survey of 97 countries worldwide in McKinsey, 40 per cent of business respondents who received more than $1 billion annually indicated that they had deployed Agent on a scale of at least one business function, up from 27 per cent in 2025.
Unlike the traditional question-and-answer, AI agent was able to break down a task into multiple steps, repeatedly calling models, using tools, implementing codes and adjusting actions based on feedback. This may not only lead to more token consumption, but also to more work on CPU than model reasoning.
According to Guo Wai, Vice-President of the Intel Marketing Group and Managing Director of the China District, the distribution relationship between CPU and GPU was still being discussed in the first half of this year at a scale of 1:8 and 1:4; in the second half of the year, the expectations of the industry for the continued growth of CPU demand became more apparent, and even suggested a possible evolution towards 1:1.
On the CPU product layout of the server, Intel identified the parallel direction of the P-core and E-core lines as early as 2023 to meet the needs of different work loads for computing performance, core density and energy efficiency. Among them, performance cores focus on computational performance and are suitable for work loads requiring high individual nuclear performance; energy efficiency cores emphasize core density and energy efficiency, and are better suited to large-scale parallel, high-input load calculation tasks.
In June this year, Intel launched the Stronger 6+ Processor (Clarwater Forest) using Intel 18A. The processor integrates up to 288 energy-efficient cores, supports 12 memory channels, and the DDR5 has a maximum transmission rate of 8,000 MT/s. Intel also disclosed that a single to a strong 6+ processor could support more than 1,000 AI agent stable operations.
However, the increase in the number of CPU cores and the increased rate of memory transmission does not mean that servers can carry more AI agents in proportion. Intel technical experts, in analysing the AI agent operating load, found that the use of CPUs was quite sudden - AI agent would produce short-lived peaks in computing demand when initiating the implementation environment and mission, but the overall utilization of CPUs was not high in terms of individual Agent operations.
This contrasts with memory. At the same time, the Intel expert noted that even if AI agent had not been given a task for the time being, its operational state could still occupy the memory. It made an on-site accounting exercise: assuming that each Agent allocated 2GB memory, and one vCPU resource operated 4 Agent, it required 8GB memory; in an example of a 256 NCPU, 12 memory channels, each with 128GB memory, the average physical memory of each vCPU was only 6GB.
This means that when cloud manufacturers wish to increase the density of AI agent deployment through re-use of resources, memory capacity may be a constraint earlier than CPU computing capacity. Intel and Tencent clouds therefore further extended this collaboration, based on a maximum of 6+, to nuclear, resource movement and memory management in the operating system, trying to translate hardware capabilities into performance enhancement in actual operations.
Save the memory, and the sandbox rises fast.
TrendForce consulting data show an increase of 53 per cent - 58 per cent in the average contractual price of the server DRAM in the second quarter of 2026. Despite the slowdown, the Agency expects that prices will continue to rise by 13-18 per cent in the third quarter.
In the face of increasing procurement costs, it is urgent for cloud manufacturers wishing to operate more Agent missions on a server to find ways to accommodate more data from existing memory.
Condensed storage of memory data that are not subject to frequent access for the time being, and decompression when required, is one of the key paths to improving the utilization of server memory. But the new problem is that while space is being saved, compression and decompression also require the consumption of computing and memory access resources, and that additional latencys may slow down operational responses, especially when large numbers of missions are simultaneously requesting compression. The Tencent cloud TencentOS Server product development team tried to compress memory using the LZ4 software programme, but found that the largest latency, which was multi-routed and decompressed, was 5 ms long and did not meet actual business needs.
To address this problem, the Tencent cloud shifted to the IAA (RAM Analytic Accelerator) embedded in Intel to the Strong CPU, performing compression and decompression tasks through dedicated hardware, reducing the use of CPU computing resources and improving the processing of latency. On this basis, the Tencent cloud team also adapted the IAA drive, simplified unnecessary call paths and tested multiple batch parallel configurations, eventually choosing the IAA-batch-8 option.
The test results show that the advantage of the IAA is not only to reduce the decompression time, but also to show high combined stability. When the combined scale increased from 32 to 256, the decompression of the LZ4 software programme rose to latency, while the latency of IAA-Batch-8 remained largely stable. In a combined pressure test closer to actual operations, for 2MB memory pages, the IAA average of latency and the maximum latency have improved by about 16 and 9 times, respectively, compared to the software programme and are at the microsecond level.
In addition to memory resources, the efficiency of sandbox creation is also an important factor affecting the high density, combined deployment and implementation of AI agent.
Sandbox is an isolated environment for AI agent to perform code, call tools and avoid interference between different tasks. The Tencent cloud CubeSandbox team noted that in part LLM intensive learning scenarios, a round of training may require the activation of thousands or even hundreds of thousands of sandboxes for model assessment and reasoning. The question of how to complete a large number of implementation environments in a short period of time, while controlling resource consumption, is a problem for cloud manufacturers.
However, there are limitations to different segregation techniques: traditional packagings are activated relatively quickly, but are usually embedded within the operating systems of shared servers; complete virtual machines have an independent kernel and require additional resources. Thus, Tencent cloud CubeSandbox uses techniques such as microvirtual machines (MicroVM), snapshots and re-use of resources, which allow multiple sandboxes to share the same basic resources and store only the changes independently, thereby reducing duplicate memory occupancy.
On this basis, the Intel engineer also optimized the nuclear implementation path of some of the operating systems involved in the sandbox creation process and introduced improvements to the OpenClaudOS kernel, a joint initiative of enterprises such as Tencent, to further reduce the large number of sandboxes and launch latency.
The field-disclosed test data show that servers carrying up to the strength of 6+ processors can support 350 sandboxes and send requests for creation, up by about 46 per cent compared with other mainstream testing platforms, while operating CubeSandbox, meeting the 200 ms service level target (SLA).
It is even more important to focus on post-loading performance. Tencent cloud measurements show that even if the server carries other tasks and the overall load rises, the sandbox capacity remains relatively stable up to 6+.
Intel does not, however, believe that the existing product is fully capable of meeting the operational needs of AI agent. On-site technical experts in Intel indicated that there was still room for further optimization between the AI agent application and the soft and hardware ecology and that the team would provide feedback on a weekly basis to the CPU design team at the back end to inform subsequent product improvements.
Cooperation between the parties will also be extended to the next generation of products. According to Tencent Senior Product Manager Shashi, Tencent Clouds continues to work with Intel to develop examples of cloud servers based on the next generation-strong processor, Diamond Rapids. In accordance with Intel's previously published product plan, Diamond Rapids will use the Intel 18A-P process, with a maximum of 256 CPU cores, supporting 16-channel memory and 128 PCIe 6.0 channels, and introducing technologies such as Advanced Performance Extension (APX) and Enhanced Advanced Matrix Extension (AMX).
Thunder Peak Network (WeChat Official Account:Ray Peak Network)
It's an original Ray Peak article, which is forbidden to reproduce without authorization. For further information, see the information below.
For your account security, please verify the mailbox.
Check your mailbox.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.