CPU paintsCPU 畫圖
Four fast cores share 576 tiles. Watch the picture grow from four places.
四個很快的核心分 576 格。看圖從四個地方長出來。
Chips and Semiconductors · Unit 4 · Lesson 10 · 晶片與半導體 單元四 第十課
How Is an AI Chip Different from a CPU?AI 晶片和一般 CPU 有什麼不同？
A CPU has a few powerful cores that do one thing after another. An AI chip has hundreds or thousands of small units that all work at the same time. In a 3D model, give both chips the same two jobs and see which one wins each time. Then find out why more workers do not always make a job faster.
CPU 有少數幾個很厲害的核心，一件事做完再做下一件。AI 晶片有幾百、幾千個小小的運算單元，全部同時工作。在 3D 模型裡，把同樣的兩種工作交給這兩種晶片，看看每一次是誰贏。再想想看：為什麼幫手變多，工作不一定變快？
Big idea · 一句話看懂
A CPU is like a few expert workers who can do any job, one step after another. An AI chip, such as a GPU, is like a huge crowd of simple workers who all do the same small job at the same time. AI needs an enormous number of simple calculations that do not depend on each other, so the crowd finishes far sooner. But for a job that must go step by step, the experts are still faster.
CPU 像少數幾位什麼都會的高手，一步接一步地做。AI 晶片（例如 GPU）像一大群只會簡單工作的幫手，大家同時做一樣的小事。AI 需要數量驚人的簡單計算，而且這些計算互不相干，所以一大群幫手快得多。可是遇到一定要一步一步來的工作，還是高手比較快。
3D Model · 3D 模型
Four Experts or a Crowd? 四位高手，還是一大群幫手？
Drag to turn the model, and scroll or pinch to zoom. Choose a chip: a CPU with 4 big cores, or a GPU with 576 small units. Then choose a job. “Paint a picture” has 576 tiles, and each tile can be done on its own. “A chain of steps” has 48 steps, and each step must wait for the one before it. Watch the board fill up and compare the times.
拖曳可以旋轉模型，滾輪或雙指可以縮放。先選晶片：有 4 個大核心的 CPU，或是有 576 個小單元的 GPU。再選工作。「畫一張圖」有 576 格，每一格都可以自己做自己的。「一串步驟」有 48 步，每一步都要等上一步做完。看板子怎麼被填滿，比一比花了多少時間。
Drag to turn · 拖曳旋轉 Scroll or pinch to zoom · 滾輪／雙指縮放
This 3D model needs WebGL, which this browser does not support. The reading and the cards below still explain everything.
這個瀏覽器不支援 WebGL，無法顯示 3D 模型；下方的課文與卡片一樣能看懂。
- Blue blocks: the CPU's 4 big cores
藍色方塊：CPU 的 4 個大核心 - Orange grid: the GPU's 576 small units
橘色格子：GPU 的 576 個小單元 - Board: dark is waiting, yellow is being worked on, colored is done
板子：深色還沒做、黃色正在做、有顏色是做完了
A teaching model with example numbers. Here a CPU core finishes 12 tiles a second and a GPU unit finishes 1; these speeds, the 4 cores, and the 576 units were chosen to make the idea easy to see. Real chips differ in many more ways, and real CPUs and GPUs work together: the CPU runs the program and hands the big shared jobs to the GPU.
這是用示例數字做的教學模型。這裡的 CPU 核心每秒做 12 格，GPU 的小單元每秒做 1 格；這些速度、4 個核心和 576 個單元，都是為了讓道理容易看懂而選的。真的晶片還有很多不同的地方，而且真的 CPU 和 GPU 是一起合作的：CPU 負責執行程式，再把可以分工的大工作交給 GPU。
Model drawn by My Culture Connect as a teaching diagram. The ideas follow Wikipedia, “Graphics processing unit” and the other sources listed at the bottom of this page. · 模型由人師教育協會自繪示意；原理依維基百科「Graphics processing unit」條目與本頁最下方列出的其他出處。
Reading · 英文閱讀
10
Unit 10: How Is an AI Chip Different from a CPU?
Every computer and phone has a CPU, which is its main processor. A CPU follows a program one instruction at a time: it fetches an instruction, works out what it means, and carries it out. It does this billions of times a second. A CPU can do almost any kind of job, and modern ones have several cores, so they can do a few things at once.
每一部電腦和手機裡都有 CPU，也就是它的主要處理器。CPU 照著程式，一次執行一個指令：把指令拿出來、弄懂它的意思、然後照著做，每秒可以做幾十億次。CPU 幾乎什麼工作都能做；現在的 CPU 有好幾個核心，所以可以同時做幾件事。
In the 1990s, video games gave computers a new kind of job. A picture on a screen is made of a huge number of pixels, and the color of every pixel must be calculated again and again. Each calculation is simple, and most of them do not depend on each other. So engineers built a different kind of chip for this work: the graphics processing unit, or GPU.
1990 年代，電玩遊戲給了電腦一種新的工作。螢幕上的畫面是由非常多的像素組成的，每一個像素的顏色都要一算再算。每一次計算都很簡單，而且大部分互不相干。於是工程師為這種工作做出另一種晶片：圖形處理器，也就是 GPU。
A GPU does not have a few big cores. It has hundreds or thousands of small calculation units, and they all work at the same time. Think of painting a long wall. One expert painter is fast, but a thousand children, each with one small square to paint, will finish first. Doing many things at the same time is called working in parallel.
GPU 沒有少數幾個大核心，而是有幾百、幾千個小小的運算單元，全部同時工作。想像要粉刷一面很長的牆：一位專業油漆師傅動作很快，可是一千個小朋友每人只刷一小格，會先刷完。同一時間做很多件事，叫做「平行」運算。
Then scientists noticed something. AI programs called neural networks need the same kind of work: an enormous number of simple multiplications and additions. In 2012 a team in Toronto trained a neural network called AlexNet on two GPUs made for games. It won a famous contest for recognizing pictures, far ahead of the others. After that, GPUs became widely used for AI.
後來科學家發現了一件事：叫做「神經網路」的 AI 程式，需要的正是同一種工作——數量驚人的簡單乘法和加法。2012 年，多倫多的一個團隊用兩張原本做來玩遊戲的 GPU，訓練出一個叫 AlexNet 的神經網路。它在一場有名的圖片辨識比賽中奪冠，遠遠領先其他隊伍。從那以後，GPU 就被廣泛用在 AI 上。
Today some chips are designed only for AI. They are called AI accelerators or neural processing units, and many phones have one inside. But a crowd of small units is not always better. Some jobs must be done in order, because each step needs the answer from the step before. For those jobs, a fast CPU core still wins.
現在有些晶片是專門為 AI 設計的，叫做 AI 加速器或神經處理器（NPU），很多手機裡就有一顆。不過，一大群小單元不一定比較好。有些工作一定要照順序做，因為每一步都要用到上一步的答案。遇到這種工作，還是速度快的 CPU 核心贏。
So a computer that runs AI uses both kinds of chips. The CPU is the manager: it runs the program and decides what to do next. The GPU or AI chip is the crowd: it takes the big jobs that can be shared and finishes them quickly. Neither one is simply better. They are good at different things, and they work as a team.
所以跑 AI 的電腦兩種晶片都用。CPU 是經理：它執行程式，決定接下來要做什麼。GPU 或 AI 晶片是那一大群幫手：把可以分工的大工作接過去，很快做完。沒有誰一定比較好。它們各有擅長的事，而且是一個團隊。
Read aloud · 課文朗讀
Questions · 閱讀理解
How does a CPU do its work?
It follows a program one instruction at a time, very quickly. It has a few powerful cores and can do almost any kind of job.
Why were GPUs first built?
For graphics. A picture has a huge number of pixels, and each one needs a simple calculation that can be done separately.
Why is a GPU good for AI but not for every job?
AI needs an enormous number of simple calculations that can be done at the same time, which suits a GPU's many small units. But some jobs must be done in order, and then only one unit can work.
Words & Phrases · 生字及片語
core(n.)核心This processor has eight cores.這個處理器有八個核心。task(n.)工作；任務Our first task is to clean the classroom.我們的第一項工作是打掃教室。parallel(adj.)平行的；同時進行的The two teams worked in parallel.兩隊同時進行工作。graphics(n.)圖形；畫面The graphics in this game are beautiful.這款遊戲的畫面很漂亮。pixel(n.)像素A photo is made of millions of pixels.一張照片是由幾百萬個像素組成的。multiply(v.)乘Multiply six by seven. What do you get?六乘以七，答案是多少？train(v.)訓練She trained her dog to sit.她訓練她的狗坐下。in order(phr.)按照順序Please read the steps in order.請按照順序讀這些步驟。
Quick Check · 小測驗
1. What is special about a GPU?
2. What were GPUs first made for?
3. What kind of work does AI need a lot of?
4. Which job is a fast CPU core still better at?
5. What does “in parallel” mean?
More Workers, More Speed? · 幫手越多就越快嗎？
The part that cannot be shared 不能分工的那一部分
Most jobs are a mix: one part can be shared among many workers, and one part must be done by a single worker, step by step. In 1967 the computer scientist Gene Amdahl showed that the part that cannot be shared sets a limit on the speed of the whole job. Move the sliders and look for the limit.
大部分的工作是混合的：一部分可以分給很多幫手，一部分只能由一個人一步一步做。1967 年，電腦科學家吉恩．阿姆達爾指出：不能分工的那一部分，決定了整件工作最快能有多快。拉拉滑桿，找找看極限在哪裡。
Or try one of these · 或試試這幾種
This is Amdahl's law: speed = 1 ÷ ((1 − p) + p ÷ n), where p is the part that can be shared and n is the number of workers. If half the work can be shared, the job can never be more than 2 times faster. If 95% can be shared, the limit is 20 times.這是阿姆達爾定律：倍數＝1 ÷（(1 − p) + p ÷ n），p 是可以分工的比例，n 是幫手的數目。如果一半的工作可以分工，整件事最多只能快 2 倍；如果 95% 可以分工，極限是 20 倍。
The whole job gets · 整件工作變成
1× as fast · 倍快
Try It in the Model · 在模型裡試試
Two chips, two jobs 兩種晶片，兩種工作
The times come from this model's example speeds. What matters is the pattern: the crowd wins when the job can be shared, and the expert wins when it cannot.
時間是用這個模型的示例速度算出來的。重要的是其中的規律：工作可以分工的時候，人多的贏；不能分工的時候，高手贏。
GPU paintsGPU 畫圖
576 slow units, one tile each, all at once. Do not blink.
576 個很慢的小單元，一人一格，同時動手。別眨眼。
CPU does the chainCPU 做一串步驟
Only one core can work, but it is quick.
只有一個核心能做，不過它很快。
GPU does the chainGPU 做一串步驟
One small unit works while 575 wait. You do not have to watch to the end.
一個小單元在做，575 個在等。不用看到最後沒關係。
Go Further · 延伸閱讀
What an AI chip needs around it AI 晶片還需要什麼
Chips and Semiconductors · Lesson 7 · 晶片與半導體 第七課What Is HBM, and Why Does AI Need It?HBM 是什麼？AI 為什麼需要它？A crowd of units needs a very wide road for its data.一大群運算單元，需要一條很寬的路來送資料。Go to the lesson · 前往這一課 →Chips and Semiconductors · Lesson 9 · 晶片與半導體 第九課Why Do Chips Get Hot?晶片為什麼會發熱？Thousands of units working at once make a lot of heat.幾千個單元同時工作，會發很多熱。Go to the lesson · 前往這一課 →
AI Chips by the Numbers · AI 晶片的數字
Six numbers about chips for AI 六個和 AI 晶片有關的數字
From the first chip called a GPU to the chip in a phone.
從第一顆被叫做 GPU 的晶片，到手機裡的那一顆。
1999the name “GPU” · 「GPU」這個名字
The GeForce 256, released on October 11, 1999, was sold as “the world's first GPU.”1999 年 10 月 11 日推出的 GeForce 256，當時用「全世界第一顆 GPU」來宣傳。
1,000sof units · 個運算單元
Modern GPUs include hundreds or thousands of calculation units.現代的 GPU 裡有幾百到幾千個運算單元。
2012AlexNet · AlexNet
A neural network trained on GPUs won a picture-recognition contest with a top-5 error of 15.3%.用 GPU 訓練的神經網路在圖片辨識比賽中奪冠，前五名錯誤率是 15.3%。
2gaming GPUs · 張遊戲用的 GPU
AlexNet was trained on two GTX 580 graphics cards for five to six days.AlexNet 用兩張 GTX 580 顯示卡訓練了五到六天。
60 millionparameters · 個參數（6,000 萬）
AlexNet had 60 million numbers to adjust and 650,000 artificial neurons.AlexNet 有 6,000 萬個要調整的數字，和 65 萬個人工神經元。
20×the limit · 極限
By Amdahl's law, if 95% of a job can be shared, it can never be more than 20 times faster.依照阿姆達爾定律，如果一件工作有 95% 可以分工，它最多只能快 20 倍。
Stories About AI Chips · AI 晶片的故事
A chip made for games, a contest in 2012, a law from 1967, and the chip in your pocket 為遊戲做的晶片、2012 年的比賽、1967 年的定律，還有口袋裡的晶片
How a chip for drawing pictures became a chip for AI.
一顆用來畫畫面的晶片，怎麼變成 AI 的晶片。
1999
Made for games為遊戲而生
A GPU is a chip designed to process images and speed up computer graphics. In 1999 Nvidia released the GeForce 256 and sold it as “the world's first GPU.” Its job was to draw game pictures quickly. Nobody was thinking about AI.
GPU 是為了處理影像、加快電腦繪圖而設計的晶片。1999 年，輝達推出 GeForce 256，用「全世界第一顆 GPU」來宣傳。它的工作是把遊戲畫面畫得又快又好，那時候沒有人想到 AI。
2012 · AlexNet
Two graphics cards兩張顯示卡
In 2012 Alex Krizhevsky, Ilya Sutskever, and their teacher Geoffrey Hinton at the University of Toronto built a neural network called AlexNet. They trained it on 1.2 million pictures using two GTX 580 graphics cards, and it took five to six days. It won that year's contest by a wide margin.
2012 年，多倫多大學的 Alex Krizhevsky、Ilya Sutskever 和他們的老師 Geoffrey Hinton 做出一個叫 AlexNet 的神經網路。他們用兩張 GTX 580 顯示卡、120 萬張圖片來訓練它，花了五到六天。它以很大的差距贏得那一年的比賽。
1967
Amdahl's law阿姆達爾定律
Long before GPUs, Gene Amdahl presented an idea at a computer conference in 1967. However many processors you add, the part of a job that cannot be shared limits the speed of the whole job. It is still used today to predict how much faster a job will run on many processors.
早在 GPU 出現之前，吉恩．阿姆達爾就在 1967 年的一場電腦研討會上提出一個想法：不管加多少個處理器，工作裡不能分工的那一部分，都會限制整件工作的速度。今天，人們還在用它預測一件工作放到很多處理器上能快多少。
NPU
Chips made only for AI專門為 AI 做的晶片
Some chips are now designed only for AI. They are called neural processing units (NPUs) or AI accelerators. Google began using its own, the Tensor Processing Unit, in 2015. AI accelerators are also built into the chips of many phones, where they can do AI work quickly or with less power.
現在有些晶片是專門為 AI 設計的，叫做神經處理器（NPU）或 AI 加速器。Google 從 2015 年開始使用自己設計的 TPU。很多手機的晶片裡也有 AI 加速器，讓 AI 的工作做得更快，或是更省電。
Myth vs. Fact · 常見迷思
Four things people get wrong 四個常見的誤會
✗ Myth · 迷思An AI chip is just a faster CPU.AI 晶片就是比較快的 CPU。
✓ Fact · 事實It is built differently: a great many small units that work at the same time, instead of a few powerful cores.它的做法不一樣：是非常多個同時工作的小單元，而不是少數幾個很強的核心。
✗ Myth · 迷思A GPU is better than a CPU at everything.GPU 做什麼都比 CPU 厲害。
✓ Fact · 事實Only at jobs that can be shared. For steps that must be done in order, a fast CPU core wins.只有在可以分工的工作上才是。遇到一定要照順序做的步驟，速度快的 CPU 核心比較快。
✗ Myth · 迷思An AI chip thinks the way a brain does.AI 晶片像大腦一樣在思考。
✓ Fact · 事實It does arithmetic: an enormous number of simple additions and multiplications, very fast.它做的是算術：數量驚人的簡單加法和乘法，而且做得非常快。
✗ Myth · 迷思Only giant computer centers have AI chips.只有巨大的電腦機房才有 AI 晶片。
✓ Fact · 事實AI accelerators are also built into the chips of many phones.很多手機的晶片裡也有 AI 加速器。
Remember It · 記憶口訣
AI chips in a sentence 一句話記住 AI 晶片
CPU: a few expertsCPU：幾位高手
Any job, one step after another.
什麼都會，一步接一步。
GPU: a crowdGPU：一大群幫手
One simple job, everybody at the same time.
一種簡單的工作，大家同時做。
Can the job be shared?這件工作能分工嗎？
If yes, the crowd wins. If no, the expert wins.
可以，人多的贏；不行，高手贏。
They are a team它們是一個團隊
The CPU manages. The GPU does the heavy shared work.
CPU 當經理，GPU 做可以分工的粗重工作。
Safety First · 安全提醒
Before you try anything 動手之前先讀
- In the classroom activities, walk, do not run, and keep scissors on the table.做課堂活動的時候用走的、不要跑，剪刀放在桌上。
- Never open a computer or a phone to look for the chips inside.不要為了找裡面的晶片而拆開電腦或手機。
- AI tools can be wrong. Check what they tell you, and ask an adult before you share personal information with one.AI 工具會出錯。它說的話要查證；要把個人資料告訴它之前，先問過大人。
Classroom Activity 1 · 課堂活動一
One Expert Against the Class 一位高手對全班
You need · 準備材料
- A grid of 30 squares drawn on the board, each with a simple sum such as 7 + 5黑板上畫 30 格，每格寫一題簡單的加法，例如 7 + 5
- The same 30 sums on small cards, one for each student, and a timer同樣的 30 題寫在小卡上，每人一張；還要一個計時器
Steps · 步驟
- Round 1: the fastest student in the class solves all 30 sums alone, one after another. Time it.第一回合：全班算得最快的同學一個人把 30 題一題一題算完。計時。
- Round 2: every student gets one card and solves it at the same moment. Time it.第二回合：每位同學拿一張卡，同一時間開始算。計時。
- Round 3: a chain. Start with 1. Each student in turn doubles the number the last student said. Time 10 steps.第三回合：一串步驟。從 1 開始，每位同學輪流把上一位說的數字乘以 2。計時 10 步。
- Talk about it: in which rounds did having more students help? In Round 3, why could only one person work at a time?討論：哪幾個回合人多有幫助？第三回合為什麼一次只能一個人做？
Round 1 is the CPU, Round 2 is the GPU, and Round 3 is the job that cannot be shared.
第一回合是 CPU，第二回合是 GPU，第三回合是不能分工的工作。
Classroom Activity 2 · 課堂活動二
Human Pixels 人體像素
You need · 準備材料
- Squared paper cut into strips, one strip for each student方格紙剪成一條一條，每人一條
- Colored pencils, and a rule on the board such as “color a square if its number can be divided by 3”色鉛筆；黑板上寫一條規則，例如「號碼可以被 3 整除的格子塗顏色」
Steps · 步驟
- Number the squares on every strip so that the whole class together covers 1 to 300.把每條方格紙編上號碼，讓全班合起來剛好是 1 到 300。
- Everyone colors their own strip by the rule at the same time. Nobody needs to ask anyone else.大家同時照規則塗自己那一條，誰都不需要問別人。
- Tape the strips together in order and look at the pattern you made.把紙條照順序貼起來，看看大家一起做出的圖案。
- Talk about it: how long would one student need for all 300 squares? Why could this job be shared so easily?討論：如果一個人塗 300 格要多久？這件工作為什麼這麼容易分工？
Each square depends only on its own number. That is exactly why pixels suit a GPU.
每一格只跟自己的號碼有關。這正是像素適合交給 GPU 的原因。
Sources · 資料出處
Facts and numbers on this page were checked against these sources (October 2026). · 本頁的事實與數字依下列資料查證（2026 年 10 月）。
- A CPU is the primary processor in a computer; its control unit directs the fetching, decoding, and execution of instructionsCPU 是電腦的主要處理器；控制單元指揮指令的擷取、解碼與執行Wikipedia, “Central processing unit” ↗
- A multi-core processor has two or more separate processing units, called cores, on one chip多核心處理器在一顆晶片上有兩個以上獨立的處理單元（核心）Wikipedia, “Multi-core processor” ↗
- A GPU is designed for digital image processing and to accelerate computer graphics; modern GPUs include hundreds or thousands of calculation units; their parallel structure led to their adoption in AIGPU 是為數位影像處理與加速電腦繪圖而設計；現代 GPU 有幾百到幾千個運算單元；平行的結構讓它被用在 AIWikipedia, “Graphics processing unit” ↗
- The GeForce 256 was released on October 11, 1999, and marketed as “the world's first GPU”GeForce 256 在 1999 年 10 月 11 日推出，以「全世界第一顆 GPU」行銷Wikipedia, “GeForce 256” ↗
- AlexNet (2012): developed by Alex Krizhevsky with Ilya Sutskever and Geoffrey Hinton at the University of Toronto; 60 million parameters and 650,000 neurons; trained on 1.2 million images for five to six days on two Nvidia GTX 580 GPUs; top-5 error of 15.3%, more than 10.8 percentage points ahead of the runner-upAlexNet（2012）：多倫多大學 Alex Krizhevsky 與 Ilya Sutskever、Geoffrey Hinton 開發；6,000 萬個參數、65 萬個神經元；用兩張 Nvidia GTX 580 GPU、120 萬張圖片訓練五到六天；前五錯誤率 15.3%，領先第二名超過 10.8 個百分點Wikipedia, “AlexNet” ↗
- A neural processing unit (NPU), also called an AI accelerator, is specialized hardware for AI; NPUs can be more efficient in speed or power use; AI accelerators are used in phone chips; their work is typically additions and multiplications; GPUs are commonly used as AI accelerators神經處理器（NPU）又稱 AI 加速器，是專為 AI 設計的硬體；在速度或耗電上可以更有效率；手機晶片裡也有 AI 加速器；它做的運算通常是加法和乘法；GPU 也常被當作 AI 加速器Wikipedia, “Neural processing unit” ↗
- The Tensor Processing Unit is a chip developed by Google for neural network machine learning; Google began using TPUs internally in 2015TPU 是 Google 為神經網路機器學習開發的晶片；Google 從 2015 年開始在內部使用Wikipedia, “Tensor Processing Unit” ↗
- Amdahl's law, named after Gene Amdahl and presented in 1967; if 50% of the work can be parallelized the best possible speedup is 2 times, and if 95% can be, 20 times阿姆達爾定律，以 Gene Amdahl 命名，1967 年發表；50% 可平行時最多快 2 倍，95% 可平行時最多快 20 倍Wikipedia, “Amdahl's law” ↗