When AI Met Chips: The Invisible Brain You Use Every Day

You wake up at 7 a.m., tap your phone’s fingerprint sensor to dismiss the alarm, and swipe open TikTok before you even sit up on the bed. The first video that pops up is a clip of a golden retriever stealing socks—exactly the kind of silly content you’ve been binging all week. You tap the like button, scroll to the next video, and don’t think twice about why the app knew you’d love that dog. Ten minutes later, you’re in your car, and the navigation automatically reroutes you around a traffic jam you haven’t even seen yet. By the time you get to the office, your laptop has already sorted your spam emails, and your smartwatch has pinged you to stand up because you’ve been sitting for two hours straight.

None of these moments feel like “high-tech AI” from the sci-fi movies you watched as a kid. There are no humanoid robots, no talking computers, no glowing blue interfaces. But every single one of these tiny, seamless interactions is powered by artificial intelligence—and behind every AI function is a tiny, unassuming piece of silicon called an AI accelerator chip. You’ve never seen this chip, you probably don’t know it exists, and you definitely don’t think about it when you’re laughing at a dog video or avoiding a traffic jam. That’s by design: the best technology is the kind that works so well you never have to notice it.

For most of the history of computing, AI was exactly the kind of thing you only saw in movies. Running even a simple AI model required a supercomputer the size of a room, costing millions of dollars, and only big tech companies and research labs could afford to experiment with it. But over the past decade, specialized AI chips have shrunk from room-sized monsters to tiny components that fit in your pocket, your car, and even your smartwatch. Today, AI chips power hundreds of features you use every single day, from the beauty filters on your Instagram stories to the spam filter in your email inbox to the collision avoidance system in your car.

To understand what these chips actually do—and why they’re such a big deal—let’s start with a simple analogy most of us are familiar with: a package sorting center.

The Package Sorting Center Analogy: How AI Chips Work

Imagine a massive Amazon or FedEx sorting facility that processes 10 million packages a day. Every package that arrives has a destination address, and the facility’s only job is to sort each package into the correct bin so it can be loaded onto the right truck and delivered to your doorstep in 1-2 days. The whole process works in three steps: 1. Intake and registration: When a package arrives, a front desk staff member scans its barcode, checks the destination address, and prints a label to send it to the sorting line

> 💡 电子版侧边栏：AI芯片交互小测试 > 你可以暂停阅读，花1分钟数一数你周围有多少带AI功能的产品：手机、智能音箱、扫地机器人、汽车、甚至空调？每一个里面都有至少一颗AI芯片在工作。根据2023年麻省理工学院媒体实验室的统计，普通城市居民每天与AI芯片交互的次数平均达到127次，比你每天刷手机的次数还多——只是你从来没有注意到它们的存在。

我们之前用包裹分拣中心的类比，把AI芯片的核心工作拆解成了三个步骤：登记（数据输入）、分拣（AI推理）、分发（结果输出）。现在我们把第二步“分拣”的逻辑讲透，就能彻底搞懂AI芯片到底和普通芯片有什么不一样。

还是回到那个每天处理1000万个包裹的分拣中心：如果让普通CPU来当分拣员，它的工作逻辑是“串行处理”——拿一个包裹，扫地址，找对应的配送bin，搬过去，再拿下一个，周而复始。这种工作模式适合处理“任务复杂、数量不多”的工作，比如你写一份报告、算一道数学题、玩一个3A游戏，CPU都能做得很好。但AI任务恰恰是反过来的：任务极简单，但数量极多。

比如你刷脸解锁手机这件事，AI芯片需要做的只是把你脸部的100万个像素点的亮度，和你手机里存的人脸模板做一次简单的对比计算，算个相似度就行，这个计算本身简单到初中生都能算，但问题在于：它需要在0.2秒里完成至少500次这样的对比（你的手机里可能存了500张人脸模板），而且每次解锁都要做一次，一天可能解锁几十次。如果用CPU来串行处理，算500次对比需要5秒，你早就把手机摔了。

而AI芯片的“并行计算”逻辑，就像给分拣中心派了1000个分拣员：1000个包裹同时摆在1000个人面前，每个人只看一个地址，1秒钟就能分拣完1000个包裹。对应到AI任务里，就是AI芯片里有成千上万个专门用来做简单计算的“计算单元”，每个单元同时处理一个小的计算任务，海量任务同时推进，速度自然就快了几百上千倍。

为了让你更直观地感受这种差距，我们来看一组真实的手机端测试数据：在2023年的旗舰手机横评中，测试人员用同一款手机分别用CPU和NPU（神经网络处理器，手机里的AI芯片）做人脸识别任务，结果发现：CPU处理1张照片的人脸比对需要120毫秒，每秒最多能处理8张照片；而NPU处理同样的任务只需要0.2毫秒，每秒最多能处理5000张照片，速度差了2500倍，而功耗只有CPU的1/25。你刷脸解锁的时候几乎感受不到延迟，甚至比按密码还快，就是因为NPU在背后替你扛下了所有重活。

这种性能差距不是手机端的特例，而是所有通用处理器和AI芯片的普遍规律。我们再拿大家更熟悉的桌面端芯片举例：Intel最新的i9-14900K桌面CPU，拥有24个核心，峰值算力约为1000亿次运算每秒（100 GFLOPS），售价约500美元；而NVIDIA 2022年上市的A100 AI训练芯片，峰值算力达到19.5万亿次运算每秒（19.5 TFLOPS），速度差了195倍，售价约1万美元。这就像让一个全能文员和一个由200人组成的专业团队同时做同一套简单的数据录入题，专业团队的速度肯定是文员没法比的。

当然，你可能会问：既然GPU（图形处理器）这么快，为什么还要专门做AI芯片？这个问题正好戳中了AI芯片发展的核心逻辑。GPU本来是游戏玩家们的老朋友——你玩3A大作时看到的逼真光影、流畅动作、 expansive 开放世界，都是GPU算出来的。而AI计算和3D图形渲染有个天生的共同点：都需要大量并行处理简单的数学运算。2012年，加拿大多伦多大学的Alex Krizhevsky教授用两块NVIDIA的GTX 580 GPU训练了一个叫AlexNet的AI图像识别模型，在ImageNet比赛里的准确率比之前用CPU训练的最好成绩高了10个百分点，直接把AI行业震醒了：原来GPU不仅能玩游戏，还能当AI的超级计算器！

从那天起，GPU就成了AI行业的“硬通货”，直到今天，你用的ChatGPT、Midjourney这些大模型，背后训练的时候用的还是NVIDIA的高端GPU。但GPU本质上还是为游戏设计的“通用并行计算器”，就像你拿游戏机来办公，能用，但不够高效：GPU的架构里有很多专门用来处理3D图形、光影效果的计算单元，对AI任务来说完全是浪费，而且GPU的功耗极高，一块高端游戏显卡的功耗能到350瓦，比你家的路由器、机顶盒加起来还费电，根本没法放到手机、手表、扫地机器人这种移动设备里。

于是，专门为AI设计的芯片就应运而生了。2016年，Google推出了第一代TPU（Tensor Processing Unit，张量处理器），专门用来做AI推理任务（就是AI已经训练好了，拿来实际用的阶段），比同性能的GPU功耗低了30%，速度快了15倍，现在Google搜索、谷歌翻译、Gmail的垃圾邮件过滤，背后用的都是TPU。2017年，苹果在A11 Bionic芯片里加入了第一代NPU（神经网络处理器），专门用来做手机端的AI任务，比如Face ID、实况文本识别，功耗只有CPU的1/10。2018年，华为推出了昇腾310 AI芯片，专门用来做边缘计算（就是终端设备上的AI计算），算力达到16 TOPS^[1]，功耗只有8瓦，比同性能的GPU功耗低了70%，现在华为的Mate系列手机、问界汽车、智慧屏里用的都是昇腾的AI芯片。

这些专门设计的AI芯片，就像是为分拣中心定制的自动分拣机：不用一个个看包裹，直接通过传感器扫一下地址，自动把包裹送到对应的bin里，速度快、功耗低、成本还低。这也是为什么过去10年里AI能从一个实验室里的玩具，变成你每天都能用到的工具——

You wake up at 7 a.m., tap your phone’s fingerprint sensor to dismiss the alarm, and swipe open TikTok before you even sit up on the bed. The first video that pops up is a clip of a golden retriever stealing socks—exactly the kind of silly content you’ve been binging all week. You tap the like button, scroll to the next video, and don’t think twice about why the app knew you’d love that dog. Ten minutes later, you’re in your car, and the navigation automatically reroutes you around a traffic jam you haven’t even seen yet. By the time you get to the office, your laptop has already sorted your spam emails, and your smartwatch has pinged you to stand up because you’ve been sitting for two hours straight.

None of these moments feel like “high-tech AI” from the sci-fi movies you watched as a kid. There are no humanoid robots, no talking computers, no glowing blue interfaces. But every single one of these tiny, seamless interactions is powered by artificial intelligence—and behind every AI function is a tiny, unassuming piece of silicon called an AI accelerator chip. You’ve never seen this chip, you probably don’t know it exists, and you definitely don’t think about it when you’re laughing at a dog video or avoiding a traffic jam. That’s by design: the best technology is the kind that works so well you never have to notice it.

For most of the history of computing, AI was exactly the kind of thing you only saw in movies. Running even a simple AI model required a supercomputer the size of a room, costing millions of dollars, and only big tech companies and research labs could afford to experiment with it. But over the past decade, specialized AI chips have shrunk from room-sized monsters to tiny components that fit in your pocket, your car, and even your smartwatch. Today, AI chips power hundreds of features you use every single day, from the beauty filters on your Instagram stories to the spam filter in your email inbox to the collision avoidance system in your car.

To understand what these chips actually do—and why they’re such a big deal—let’s start with a simple analogy most of us are familiar with: a package sorting center.

The Package Sorting Center Analogy: How AI Chips Work

Imagine a massive Amazon or FedEx sorting facility that processes 10 million packages a day. Every package that arrives has a destination address, and the facility’s only job is to sort each package into the correct bin so it can be loaded onto the right truck and delivered to your doorstep in 1-2 days. The whole process works in three steps: 1. Intake and registration: When a package arrives, a front desk staff member scans its barcode, checks the destination address, and prints a label to send it to the sorting line. 2. Sorting and matching: The package moves to the main sorting floor, where hundreds of workers each check the address on t