# 计算 · 日跟踪 · 2026-10-06

> 日期：2026-10-06（Asia/Shanghai，周二）。专题正文：`intel/topics/compute.md`。原料：`/workspace/intel/archive/2026-10-06/`（窗口 10-05 早报后→10-06 早）；对照 2026-10-05 日跟踪。抓取：主跑 Saved 267 / skipped 28 / errors 40（+ Scale 源修复补抓 14）；blocked_hosts：blocksandfiles.com、docs.redhat.com、e.huawei.com、support.huawei.com、www.nextplatform.com、www.reddit.com、www.theregister.com。

## 摘要
- **勘误（重要）· Xeon 6+ Clearwater Forest 不支持 MRDIMM**：STH 修订原文为 "12 channels of DDR5 (not MRDIMM)"，每核带宽 2.67（288 核）–5.34（144 核）GB/s 按普通 DDR5-8000 估算——**10-05 专题/日报「含 MRDIMM」作废**；其余（4 SKU、144–288 E-core、低 TDP 1.14–1.15 W/core、list $4.6K→~$15K）不变。
- **厂商自测 · VMware AI Part 3**：VCF 9.1、Dell XE7745（2×EPYC 9555、256 逻辑核、768GB、4×RTX 6000 Blackwell 96GB）上 Llama3.1 8B 推理——vCPU 16→128、内存 32→512GB "little or no impact"；25%–50% 核即与裸机持平；Llama3 70B（4×H200，1 vCPU）与 HammerDB（12 vCPU、1.8M TPM）同机互不影响——**非第三方基准**；规划含义：GPU 节点 CPU/内存可超配给其他负载的说法需我方自测验证。
- 轻 · XCP-ng 新增 VM 选项缓解部分 **AMD GPU 直通**问题（后续拟自动开启）。
- 续档 · EPYC 9006 Venice × Agentic；TRX50。

## 原文摘录

### 勘误 · STH Intel Xeon 6+ SKU 分析（原文修订）
- 归档：`industry/intel-xeon-6-sku-list-and-value-analysis-clearwater-forest-runs-wide-0d5dd057.txt`
- URL：https://www.servethehome.com/intel-xeon-6-sku-list-and-value-analysis-clearwater-forest-runs-wide/

> … customers get it all: 2 socket support, 12 channels of DDR5 (not MRDIMM) memory, 96 PCIe lanes, and Intel's new Application Energy Telemetry (AET)…  
> … assuming 12 channels of DDR5-8000 (not MRDIMMs), this spans from 2.67GB/second of memory bandwidth per core for the 288-core SKU to 5.34GB/second for the 144-core SKU.  

### VMware · Optimizing AI Deployments with VCF Part 3（厂商自测）
- 归档：`vmware/optimizing-ai-deployments-with-vmware-cloud-foundation-part-3-f4be665b.txt`
- URL：https://blogs.vmware.com/cloud-foundation/2026/10/05/optimizing-ai-deployments-with-vmware-cloud-foundation-part-3/

> Even when we allocated only 25%-50% of available CPU cores to the VM, the inference performance is on par with bare metal.  
> … changing the number of vCPUs while fixing the memory size, and changing the memory size while fixing the number of vCPUs, had little or no impact on the performance of AI workloads that use GPUs for acceleration.  
> The HammerDB workload runs at 1.8 million TPM … The database server VM runs at 75% utilization on 12 vCPUs. The LLM-inference uses 1 vCPU. The results show that neither workload experiences a measurable performance impact.