过去三年，AI每进化一次，算力重心就搬一次家。训练时代，GPU是绝对主角；到了Agentic AI时代，一件反直觉的事正在发生——真正决定AI干活快慢的，可能是一排排不起眼的CPU。云尖信息C1000e X7高密算力液冷整机柜，正是为这个变化而来。Over the past three years, each new wave of AI has shifted the center of gravity in computing. During the model-training era, GPUs took center stage. Now, with the rise of Agentic AI, a less obvious shift is taking place: rows of unassuming CPUs may increasingly determine how quickly AI systems get work done. CNIT’s C1000e X7 high-density liquid-cooled rack system is designed for this transition.

此次Intel Xeon 6 AP（Granite Rapids-AP）平台，40个双路液冷计算节点不变，采用契合Agentic AI的处理器。The latest configuration adopts Intel® Xeon® 6 AP (Granite Rapids-AP) processors while retaining 40 dual-socket, liquid-cooled compute nodes, delivering a CPU platform tailored to Agentic AI workloads.

一、Agentic AI 时代，CPU 重新站上 C 位 Agentic AI Puts CPUs Back in the Spotlight

传统生成式 AI 的工作流很“直”：输入、前向传播、输出，GPU 扛下几乎全部重计算，CPU 只负责数据预处理和 I/O 路由，GPU 与 CPU 的配比通常在 1:8 。Traditional generative AI follows a relatively linear workflow: input, forward pass, and output. GPUs handle nearly all compute-intensive operations, while CPUs primarily perform data preprocessing and I/O routing. The typical GPU-to-CPU ratio is 1:8.

Agentic AI完全不同。一个智能体完成任务，要在“任务规划 → 多智能体协同 → 工具调用 → 代码沙箱执行 → 验证重试”的循环里反复流转。这个循环里，几乎每一步都跑在CPU上：Agentic AI works differently. To complete a task, an agent repeatedly cycles through task planning→multi-agent coordination→tool invocation→code sandbox execution→validation and retry. Nearlyevery stage of this loop runs on the CPU:

CPU工作总占比可达50%–90%；CPU-related work can account for 50%–90% of total workload activity;
处理时长占总时长的80%–90%；CPU-side processing can account for 80%–90% of total execution time;
GPU与CPU的合理配比，从1:8被拉高到1:2。The recommended GPU-to-CPU ratio shifts from 1:8 to 1:2.

一句话：模型越会“干活”，CPU就越忙。可多数机房里的CPU算力，还是按“配角”标准配置的——这就是新的算力缺口。The more work AI models can carry out autonomously, the busier CPUs become. Yet most data centers still provision CPU capacity as a supporting resource. This mismatch is creating a new compute gap.

二、C1000eX7：把 80 颗 Granite Rapids-AP 装进一个机柜 C1000e X7: 80 Granite Rapids-AP CPUs in a Single Rack

C1000e 的解题思路直接而极致：在 1U 高度内部署双路 CPU，用 40 个节点把算力堆到天花板。The C1000e takes a straightforward, high-density approach: two CPUs in each 1U compute node, with 40 nodes delivering substantial aggregate processing capacity.

单节点：1U空间集成双路Intel GNR AP处理器，液冷散热模组、主板、机箱全套齐备；Per node: a complete 1U system integrating two Intel® GNR-AP processors, liquid-cooling modules, a motherboard, and a chassis.

单节点配置：Single-Node Specifications:

单机柜：47U满配40个1U液冷节点，最高80颗CPU、1万+核心；Per rack: one fully populated 47U rack houses 40 liquid-cooled 1U nodes, with up to 80 CPUs and over 10,000 CPU cores.
集群：400G RDMA高速网络，多链路灵活扩展，40节点高效协同。Cluster: high-speed 400G RDMA networking and flexible multi-link expansion enable efficient collaboration across all 40 nodes.

Granite Rapids-AP（Intel Xeon 6 AP 平台）有几项天然优势：x86 生态最完整，Agent 开发框架、工具链、沙箱运行时几乎零迁移；芯片内置 IAA、DSA、QAT、AMX 等硬件加速器，专门卸掉 Agent 场景里的“重体力活”；还有覆盖六大区域 52 项的 RAS 能力，为 7×24 高并发运行兜底。The Granite Rapids-AP platform (Intel® Xeon® 6 AP) brings several inherent advantages. Its mature x86 ecosystem enables Agent development frameworks, toolchains, and sandbox runtimes to migrate with minimal effort. Integrated Intel® IAA, DSA, QAT, and AMX capabilities help offload demanding operations in Agentic AI workloads. The platform also provides 52 RAS capabilities across six categories to support reliable 24/7 operation under high concurrency.

散热上采用风液混合方案——冷板式液冷带走约70%的热量，剩余30%由风冷辅助排出，整机PUE控制在1.2，在密度与能效之间找到平衡。A hybrid air-and-liquid cooling design removes approximately 70% of system heat through direct-to-chip cold plates, with the remaining 30% handled by supplemental air cooling. The system targets an overall PUE of 1.2, balancing computing density and energy efficiency.

三、三大优势 Three Key Advantages

1、完全解耦：想怎么装都行 Fully Decoupled: Flexible Configuration and Deployment

C1000e摒弃了传统“机柜与节点绑定交付”的固化模式，把每个计算节点做成独立的1U液冷单元，即插即用：
快插连接：冷却液走快插流体连接器，支持上万次插拔零泄漏，供电走CPRS集中电源+电源线，拆装不用停柜；Quick-disconnect connections: coolant flows through quick-disconnect fluid couplings designed for more than 10,000 leak-free mating cycles. Power is delivered through a CPRS centralized power system and power cables, allowing nodes to be removed or installed without shutting down the entire rack.

部署自由：单节点可独立上架、独立组网，适配边缘站点和老旧机房局部扩容；也可批量整柜交付，快速搭起一套HPC超算；Flexible deployment: each node can be rack-mounted and networked independently, making it suitable for edge sites or incremental expansion of existing data centers. Alternatively, complete racks can be delivered in bulk to deploy an HPC cluster rapidly.

平滑升级：节点与机柜彻底解耦，未来CPU平台升级换代时，只需更换节点内的CPU与主板，机箱结构、机柜、液冷管网、供电网络均可复用，保护机柜级长期投资。Seamless upgrades: compute nodes are decoupled from the rack. During future CPU platform upgrades, only the CPUs and motherboards need to be replaced; the chassis, racks, liquid-cooling pipework, and power distribution infrastructure can be reused, protecting long-term rack-level investment.

2、极致性能：为 Agent 沙箱而生 High-Density Performance: Built for Agent Sandboxes

AgenticAI 最大的算力消耗，不在 GPU 推理，而在海量的代码沙箱。智能体每调用一次工具、跑一段代码、验证一次结果，都要拉起一个隔离沙箱。沙箱拉得越快、装得越多，Agent的并发能力就越强。In Agentic AI, substantial computing demand comes not only from GPU inference but also from large numbers of code sandboxes. Each time an agent invokes a tool, executes code, or validates a result, it may need to launch an isolated sandbox. Faster sandbox startup and higher sandbox density translate into greater agent concurrency.

Intel实测显示，单个Agent沙箱稳定运行约需0.25 vCPU，即1颗核心可支撑约4个沙箱（1:4超分）。照此口径，C1000e单柜1万+核心，理论上可同时拉起4万级的Agent沙箱。Intel testing indicates that a single Agent sandbox requires approximately 0.25 vCPU for stable operation. By this measure, one CPU core can support around four sandboxes (a 1:4 core-to-sandbox ratio). With more than 10,000 cores per rack, the C1000e could theoretically host on the order of 40,000 concurrent Agent sandboxes.

密度拉满后，瓶颈会转移到内存。Xeon 6的IAA、DSA硬件加速器正是为此而来：IAA负责沙箱内存的压缩与超分，DSA负责海量相同内存页的去重，把本该CPU干的“体力活”卸载到专用硬件，让更多核心留给真正的业务计算。At this scale, memory can become the next bottleneck. Intel® Xeon® 6 incorporates IAA and DSA hardware accelerators to address this challenge: IAA handles sandbox-memory compression and overcommit, while DSA handles the deduplication of large numbers of identical memory pages. Moving these tasks to dedicated hardware leaves more CPU cores available for application workloads.

3、运维方便：算力永续在线Streamlined Operations: Keep Computing Resources Online

算力设备买回来只是开始，运维体验决定它能不能一直创造价值。C1000e做了三层设计：Purchasing computing hardware is only the beginning. Day-to-day operability determines whether that investment delivers lasting value. The C1000e addresses operations and maintenance in three ways:

分钟级上线：整机工厂预制、现场快速部署，业务分钟级就绪；节点快插拆装，扩容升级不停柜；Ready in minutes: factory-integrated systems support rapid on-site deployment and workload readiness within minutes. Quick-disconnect nodes allow expansion and upgrades without shutting down the rack.
液冷安全防护：微量漏液检测+全链路防护，液冷链路全程监测、自动预警；Liquid-cooling protection: trace-leak detection and end-to-end safeguards continuously monitor the liquid-cooling circuit and trigger automated alerts.
智能管控：机柜+节点两级管理，温度、流量、功耗核心参数实时可视；全链路自动化预警、故障隔离与恢复，单节点异常不影响整体集群；双路供电、双泵热备、N+1冗余，核心部件热插拔。Intelligent management: two-level rack-and-node management provides real-time visibility into temperature, coolant flow, and power consumption. Automated alerts, fault isolation, and recovery help prevent a single-node failure from affecting the overall cluster. Dual power feeds, redundant pumps with hot standby, N+1 redundancy, and hot-swappable critical components further improve availability.

四、它适合谁？ Who Is It For?

做AI Agent的团队——为多智能体协同、工具调用、代码沙箱提供CPU编排底座，与GPU集群按1:2配比组网；AI Agent development teams: a CPU-based orchestration platform for multi-agent collaboration, tool invocation, and code sandboxes, with CPU resources networked alongside GPU clusters at a 1:2 GPU-to-CPU ratio.
跑HPC任务的科研与工业用户——气象预报、工业仿真、生物医药，算力节点灵活增减；Research and industrial HPC users: flexible compute capacity for weather forecasting, industrial simulation, and biomedical workloads, with nodes added or removed as requirements change.

算力市场的竞争，正在从“单芯片参数”转向“集群落地效率”。当Agentic AI把CPU重新推回舞台中央，一台为CPU而生的整机柜，就是这场变局里最直接的答案。而这只是开始——C1000e产品家族即将带来单柜240颗CPU的更高密度版本，敬请期待！Competition in computing infrastructure is shifting from individual-chip specifications toward the efficiency of deploying complete clusters. As Agentic AI brings CPUs back into focus, a rack system designed around CPU density offers a direct response to this shift. And there is more to come: the C1000e family will soon introduce an even higher-density configuration supporting up to 240 CPUs per rack. Stay tuned.