過去三年，AI每進化一次，算力重心就搬一次家。訓練時代，GPU是絕對主角；到了Agentic AI時代，一件反直覺的事正在發生——真正決定AI干活快慢的，可能是一排排不起眼的CPU。云尖信息C1000e X7高密算力液冷整機柜，正是為這個變化而來。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）平臺，40個雙路液冷計算節點不變，采用契合Agentic AI的處理器。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 時代，CPU 重新站上 C 位 Agentic AI Puts CPUs Back in the Spotlight

傳統生成式 AI 的工作流很“直”：輸入、前向傳播、輸出，GPU 扛下幾乎全部重計算，CPU 只負責數據預處理和 I/O 路由，GPU 與 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。一個智能體完成任務，要在“任務規劃 → 多智能體協同 → 工具調用 → 代碼沙箱執行 → 驗證重試”的循環里反復流轉。這個循環里，幾乎每一步都跑在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工作總占比可達50%–90%；CPU-related work can account for 50%–90% of total workload activity;
處理時長占總時長的80%–90%；CPU-side processing can account for 80%–90% of total execution time;
GPU與CPU的合理配比，從1:8被拉高到1:2。The recommended GPU-to-CPU ratio shifts from 1:8 to 1:2.

一句話：模型越會“干活”，CPU就越忙。可多數機房里的CPU算力，還是按“配角”標準配置的——這就是新的算力缺口。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 顆 Granite Rapids-AP 裝進一個機柜 C1000e X7: 80 Granite Rapids-AP CPUs in a Single Rack

C1000e 的解題思路直接而極致：在 1U 高度內部署雙路 CPU，用 40 個節點把算力堆到天花板。The C1000e takes a straightforward, high-density approach: two CPUs in each 1U compute node, with 40 nodes delivering substantial aggregate processing capacity.

單節點：1U空間集成雙路Intel GNR AP處理器，液冷散熱模組、主板、機箱全套齊備；Per node: a complete 1U system integrating two Intel® GNR-AP processors, liquid-cooling modules, a motherboard, and a chassis.

單節點配置：Single-Node Specifications:

單機柜：47U滿配40個1U液冷節點，最高80顆CPU、1萬+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高速網絡，多鏈路靈活擴展，40節點高效協同。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 平臺）有幾項天然優勢：x86 生態完整，Agent 開發框架、工具鏈、沙箱運行時幾乎零遷移；芯片內置 IAA、DSA、QAT、AMX 等硬件加速器，專門卸掉 Agent 場景里的“重體力活”；還有覆蓋六大區域 52 項的 RAS 能力，為 7×24 高并發運行兜底。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.

散熱上采用風液混合方案——冷板式液冷帶走約70%的熱量，剩余30%由風冷輔助排出，整機PUE控制在1.2，在密度與能效之間找到平衡。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、三大優勢 Three Key Advantages

1、完全解耦：想怎么裝都行 Fully Decoupled: Flexible Configuration and Deployment

C1000e摒棄了傳統“機柜與節點綁定交付”的固化模式，把每個計算節點做成獨立的1U液冷單元，即插即用：
快插連接：冷卻液走快插流體連接器，支持上萬次插拔零泄漏，供電走CPRS集中電源+電源線，拆裝不用停柜；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：單節點可獨立上架、獨立組網，適配邊緣站點和老舊機房局部擴容；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.

平滑升級：節點與機柜徹底解耦，未來CPU平臺升級換代時，只需更換節點內的CPU與主板，機箱結構、機柜、液冷管網、供電網絡均可復用，保護機柜級長期投資。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、極致性能：為 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，而在海量的代碼沙箱。智能體每調用一次工具、跑一段代碼、驗證一次結果，都要拉起一個隔離沙箱。沙箱拉得越快、裝得越多，Agent的并發能力就越強。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實測顯示，單個Agent沙箱穩定運行約需0.25 vCPU，即1顆核心可支撐約4個沙箱（1:4超分）。照此口徑，C1000e單柜1萬+核心，理論上可同時拉起4萬級的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.

密度拉滿后，瓶頸會轉移到內存。Xeon 6的IAA、DSA硬件加速器正是為此而來：IAA負責沙箱內存的壓縮與超分，DSA負責海量相同內存頁的去重，把本該CPU干的“體力活”卸載到專用硬件，讓更多核心留給真正的業務計算。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、運維方便：算力永續在線Streamlined Operations: Keep Computing Resources Online

算力設備買回來只是開始，運維體驗決定它能不能一直創造價值。C1000e做了三層設計：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:

分鐘級上線：整機工廠預制、現場快速部署，業務分鐘級就緒；節點快插拆裝，擴容升級不停柜；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.
液冷安全防護：微量漏液檢測+全鏈路防護，液冷鏈路全程監測、自動預警；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：機柜+節點兩級管理，溫度、流量、功耗核心參數實時可視；全鏈路自動化預警、故障隔離與恢復，單節點異常不影響整體集群；雙路供電、雙泵熱備、N+1冗余，核心部件熱插拔。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、它適合誰？ Who Is It For?

做AI Agent的團隊——為多智能體協同、工具調用、代碼沙箱提供CPU編排底座，與GPU集群按1:2配比組網；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任務的科研與工業用戶——氣象預報、工業仿真、生物醫藥，算力節點靈活增減；Research and industrial HPC users: flexible compute capacity for weather forecasting, industrial simulation, and biomedical workloads, with nodes added or removed as requirements change.

算力市場的競爭，正在從“單芯片參數”轉向“集群落地效率”。當Agentic AI把CPU重新推回舞臺中央，一臺為CPU而生的整機柜，就是這場變局里最直接的答案。而這只是開始——C1000e產品家族即將帶來單柜240顆CPU的更高密度版本，敬請期待！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.