# 端侧推理 · 日跟踪 · 2026-09-14

> 日期：2026-09-14（Asia/Shanghai）。专题正文：`intel/topics/edge-inference.md`。原料：`/workspace/intel/archive/2026-09-14/`；对照 2026-09-11 日跟踪。

## 摘要

- **新稿交叉 · Google agentic infra 调研：** **90%** 受访者认为在边缘部署生成式模型对本组织重要（72% 极重要/很重要）；理由框架：时延、成本、断网韧性——集中训练/复杂任务 vs 边缘选定推理。（**调研/研究推算**）
- **续档再抓 · Qualcomm Hexagon NPU：** Element Accelerator、32K 上下文、KV-cache、MoE ~30B/3B、INT4 prefill 宣称至 +50%（**厂商自称**）——手机/端侧 SoC，**非门店 HCI 推理一体机**。
- **续档轻 · Arm Neoverse CSS N4：** N 系列常见于 DPU/网络器具。
- **本日无公开战略级「边缘 AI HCI 可售一体机」新通稿。**

## 原文摘录

### Google · 边缘 90%

- 归档：`industry/agentic-ai-pushes-enterprise-infrastructure-toward-an-upgrade-cycle-go-6af11bcd.txt`  
  > Ninety percent of respondents said deploying generative AI models at the edge is important to their organizations, with 72% rating it extremely or very important.  
  > … latency, cost and operational resilience… Local inference can also reduce reliance on central cloud compute… while edge deployments can continue operating when the primary network connection is interrupted.

### Qualcomm Hexagon · 续

- `industry/qualcomm-talks-next-gen-oryon-cpu-adreno-gpu-and-hexagon-npu-a88def6a.txt`  
  > The NPU features a new Element Accelerator… context lengths up to 32K… KV-cache acceleration.  
  > Mixture of Experts (MoE) models up to 30 billion total parameters… roughly 3 billion active parameters…  
  > Inference workloads are dispatched to the GPU or NPU, with a cloud path available for offloading.

## 检索范围

- `archive/2026-09-14/industry`（Google、Qualcomm、Arm）；对照 2026-09-11 edge-inference