# 端侧推理专题日跟踪 · 2026-09-11

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

## 摘要

- **新稿交叉 · Qualcomm Hexagon NPU：** Element Accelerator、32K 上下文、KV-cache 加速、MoE 约 30B/3B、INT4 prefill 宣称至 +50%（**厂商自称**）——**手机/端侧 SoC 叙事，不是门店 HCI 推理 SKU**。
- **续档轻交叉 · Arm Neoverse CSS N4：** N 系列常见于 DPU/网络器具——边缘加速子系统 IP。
- **本日无公开战略级「边缘 AI HCI 一体机」可售新稿。**

## 原文摘录

### Qualcomm Hexagon · 端侧

- 归档：`industry/qualcomm-talks-next-gen-oryon-cpu-adreno-gpu-and-hexagon-npu-a88def6a.txt`  
  > The NPU features a new Element Accelerator for transformer operations… supports context lengths up to 32K and includes KV-cache acceleration.  
  > Mixture of Experts (MoE) models up to 30 billion total parameters run on the Hexagon NPU, with roughly 3 billion active parameters…  
  > The Sensing Hub captures voice input and routes it to the Oryon CPU… Inference workloads are dispatched to the GPU or NPU, with a cloud path available for offloading.

## 检索范围

- archive/2026-09-11/industry（Qualcomm、Arm）；对照 2026-09-10 edge-inference 日跟踪