一颗 CPU 里藏着三块算力：CPU / 核显 / NPU 到底各自擅长什么

如果你到现在还觉得'CPU 里只有 CPU'，那这颗 Core Ultra 5 235H 会刷新认知。它一片 SoC 里塞了三块独立的'算力'：14 核的 CPU、Arc 140T 核显（共享 12GB 显存）、还有一块 Intel AI Boost NPU。三块硬件各吃各的电、各干各的活——而且，可以同时跑。这篇把它们的性格讲清楚，你就知道自己的任务该交给谁。

基于一台实测机器（Core Ultra 5 235H / Arc 140T 12GB共享 / Intel AI Boost NPU / OpenVINO 2026.5.0.dev）和一整天的 YOLOv8n 推理实测。官方规格（核数/TOPS/TDP）据 Intel ARK / Wikipedia，实测架构参数 2026-09-12。

这颗 235H 的官方规格

项	官方值
CPU 核心	14 核 14 线程（4 P + 8 E + 2 LP-E，无超线程）
频率	P 核 2.4/5.0 GHz，E 核 1.8 GHz
L3 缓存	18 MB
核显	Arc 140T，8 Xe 核心，~2350 MHz
NPU	Intel AI Boost，官方 13 TOPS
TDP	28W base / 115W 峰值
发布	2025 年 1 月（Arrow Lake-H）

注意一个容易错的点：235H 是 14 核（4P+8E+2LP-E），不是 6P 的 14 核——285H 那种旗舰才有 6P。而三块算力里只有 CPU 是'万能'的（所以没有独显的机器，大模型最后一般在核显上跑，见本站核显跑 Gemma4 实测）。真正容易被误会的，是 NPU 那 13 TOPS 的量级到底「能不能当显卡用」——答案是不能，但它有自己的主场。

可以粗暴地类比成一个团队：CPU 是项目经理兼杂活工，核显是搬砖壮劳力，NPU 是专业技能岗。

引擎	擅长	不擅长	能耗姿态
CPU（P/E 核）	通用逻辑、串行分支、调度，什么都能干	重矩阵运算很费劲	待机低、负载上涨快
Arc 核显 iGPU	并行吞吐、矩阵乘法、批量任务	串行逻辑、精确时序	高负载高效，低负载也耗电
NPU	低功耗持续 AI 推理、专用算子	通用编程是不可能的	始终低功耗

OpenVINO 眼中的三块设备

应用不是自己挑硬件跑的，而是通过运行时。OpenVINO 把 CPU/GPU/NPU 三个设备暴露给开发者的就是三个名字：CPU、GPU、intel:NPU。同一份 YOLOv8n 模型，切到哪个设备就在哪跑：

from ultralytics import YOLO

model = YOLO("yolov8n_openvino_model")
for device in ["cpu", "gpu", "intel:NPU"]:
    r = model.predict(source="f.jpg", device=device, classes=[0])

同一模型，三块硬件实测

用 YOLOv8n·OpenVINO 在同一台 235H 上实测（抽帧路径一致），最能说明性格差异：核显追求峰值吞吐必须多进程喂，NPU 胜在小而稳定。

设备	最优吞吐（张/s）	实际规律
NPU ×3 进程	72	天花板低但稳定低功耗
GPU ×4 进程	147.8	多进程才喂得饱，峰值更高
混合 GPU×4 + NPU×3	240	两芯并行，各吃各的

NPU+核显同时开工没有任何调度冲突，加起来还能更快（240 张/s ≈ 核显的 1.6 倍、NPU 的 3.3 倍）。这就是'三块算力可并行'最直观的证据：不是谁替代谁，而是叠着用。完整吞吐曲线见三档吞吐天梯。

任务该交给谁：决策表

任务	首选	为什么
长时间监控逐秒检测	NPU	低功耗持续跑，不抢 CPU
海量批量抽帧推理	核显（多进程）	峰值吞吐更高
要最高吞吐	核显+NPU 混合	两芯叠加不冲突
LLM 文本生成	核显（Vulkan/OpenVINO）	NPU 只专精 CNN 类算子，不适合大模型
临时小脚本推理	CPU	零初始化，启动即跑

一句话总结

核显负责猛，NPU 负责省，CPU 负责兜底。没有独显的机器，把 '持续、重复、固定形状' 的推理交给 NPU，把 '大批量、要峰值' 的推理交给多进程核显，两个一起上是白捡的吞吐。唯一不能指望的，是让 NPU 去跑 LLM——那是核显的活。

延伸阅读：NPU 上怎么把 YOLOv8n 跑起来 → OpenVINO 导出 + 官方路径教程；踩过的坑 → NPU OpenVINO 避坑录；核显跑大模型全数据 → 核显跑 Gemma4 实测。

If you still think 'the CPU contains only a CPU', the Core Ultra 5 235H will reset that idea. This one SoC packs three independent compute engines: a 14-core CPU, the Arc 140T iGPU (12GB shared) and an Intel AI Boost NPU. Three chips, separate power budgets, separate jobs — and, crucially, they can run at the same time. Learn each engine's personality here, and you'll know who to hand each job to.

Based on one measured box (Core Ultra 5 235H / Arc 140T, 12GB shared / Intel AI Boost NPU / OpenVINO 2026.5.0.dev) and a full day of YOLOv8n inference tests. Official specs (cores/TOPS/TDP) per Intel ARK / Wikipedia; measured numbers 2026-09-12.

Official specs of this 235H

Item	Official
CPU cores	14 cores / 14 threads (4 P + 8 E + 2 LP-E, no SMT)
Clocks	P 2.4/5.0 GHz, E 1.8 GHz
Smart Cache	18 MB
iGPU	Arc 140T, 8 Xe cores, ~2350 MHz
NPU	Intel AI Boost, 13 TOPS
TDP	28W base / 115W peak
Released	Jan 2025 (Arrow Lake-H)

Don't get this wrong: the 235H packs 14 cores as 4P+8E+2LP-E, not 6P — the 6P flagship is the 285H class. And of the three engines, only the CPU is 'universal' (that's why, with no dGPU, LLMs end up on the iGPU — see Gemma 4 on an iGPU). The real misconception is whether the NPU's ~13 TOPS can 'act as a GPU' — it can't, but it owns a lane of its own.

A crude analogy: the CPU is the project manager + all-rounder, the iGPU is the brute-force mover, the NPU is the specialist.

Engine	Good at	Bad at	Power profile
CPU (P/E cores)	general logic, branches, scheduling — does everything	heavy matmul struggles	low idle, spikes under load
Arc iGPU	parallel throughput, matmul, batching	serial logic, precise timing	efficient sustained, costly at idle
NPU	low-power sustained AI inference, dedicated ops	can't do general compute	always low-power

Three devices as OpenVINO sees them

Apps don't pick hardware themselves — a runtime does. OpenVINO exposes the three chips to developers as three names: CPU, GPU, intel:NPU. The same YOLOv8n model runs wherever you point it:

from ultralytics import YOLO

model = YOLO("yolov8n_openvino_model")
for device in ["cpu", "gpu", "intel:NPU"]:
    r = model.predict(source="f.jpg", device=device, classes=[0])

One model, three engines, measured

The personality gap shows best on the same 235H with YOLOv8n·OpenVINO (same frame-sampling path): the iGPU wants multiple processes to peak, the NPU wins on being small and steady.

Device	Best (FPS)	Pattern
NPU ×3 进程	72	low ceiling, steady & low-power
GPU ×4 进程	147.8	needs multi-process, higher peak
hybrid GPU×4 + NPU×3	240	parallel, no contention

NPU + iGPU in parallel shows zero contention and adds up (240 FPS ≈ 1.6× the iGPU alone, 3.3× the NPU). It's the clearest proof that the three engines stack, not replace. Full curves: the three-lane ladder.

Which engine for which job: the decision table

Task	First choice	Why
sustained per-second detection	NPU	low-power sustain, frees CPU
bulk frame batch inference	iGPU (multi-process)	higher peak throughput
max throughput	iGPU + NPU hybrid	stack, no contention
LLM text generation	iGPU (Vulkan/OpenVINO)	NPU specializes in CNN ops, not LLMs
one-off small inference	CPU	zero init, run on demand

TL;DR

The iGPU is for brute force, the NPU for efficiency, the CPU for everything else. On a no-dGPU box, hand 'sustained, repeated, fixed-shape' inference to the NPU, 'bulk, peak' inference to a multi-process iGPU, and running both together is free throughput. The one thing not to expect is an NPU running an LLM — that's the iGPU's job.

Further reading:How to run YOLOv8n on an NPU → OpenVINO export + official path; the traps → NPU OpenVINO pitfalls; full iGPU LLM numbers → Gemma 4 iGPU tests.

返回文章列表