Ordinary memory一般記憶體
One lane, a long way. Watch how few cars reach the logic chip.
一個車道,路很長。看看到得了運算晶片的車有多少。
Chips and Semiconductors · Unit 3 · Lesson 7 · 晶片與半導體 單元三 第七課
An AI chip can calculate very fast, but it often has to wait for its data. HBM is memory stacked like a tower right beside the chip and joined to it by a road with more than a thousand lanes. In 3D, compare ordinary memory with four generations of HBM, then work out how many movies it could move in one second.
AI 晶片算得很快,卻常常要等資料。HBM 是像高塔一樣疊起來的記憶體,就放在晶片旁邊,中間是一條有上千個車道的路。在 3D 裡比較一般記憶體和四代 HBM,再算算它一秒可以搬幾部電影。
A fast chip is no use if its data arrives slowly. HBM, short for High Bandwidth Memory, stacks memory chips into a tower, connects the floors with tiny elevators called TSVs, and sets the tower right next to the logic chip. More than a thousand data lines join them, like a very wide, very short highway, so huge amounts of data move at once.
資料送得慢,晶片再快也沒用。HBM 是「高頻寬記憶體」(High Bandwidth Memory)的縮寫:把記憶體晶片疊成一座塔,用叫做矽穿孔的小電梯連接每一層,再把塔緊貼著運算晶片放。上千條資料線把兩邊連起來,像一條又寬又短的高速公路,一次就能送很多資料。
3D Model · 3D 模型
Drag to turn the model, and scroll or pinch to zoom. Start with “Ordinary memory”: the memory chips are far across the board, and the road has few lanes. Then choose HBM. The memory becomes a tower beside the blue logic chip, and the road becomes short and wide. Pick a generation on the right to see the lanes, the floors, and the speed change.
拖曳可以旋轉模型,滾輪或雙指可以縮放。先看「一般記憶體」:記憶體晶片在板子的另一頭,路上沒幾個車道。再選 HBM:記憶體變成藍色運算晶片旁邊的一座塔,路變得又短又寬。在右邊選世代,看車道、樓層和速度怎麼變。
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。
A teaching diagram, not to scale. The numbers of data lines, floors, and gigabytes per second come from the published HBM standards. The road, the cars, and the distances are only pictures: real data lines are thin wires inside the interposer, and data is electrical signals, not cars. The ordinary memory shown is graphics memory, whose chips each have 32 data lines; it is drawn with a single lane.
這是教學用的示意圖,不是真實比例。資料線數、樓層數和每秒幾 GB 來自公開的 HBM 標準。路、車子和距離都只是比喻:真的資料線是中介層裡的細線,資料是電的訊號,不是車子。圖裡的一般記憶體是顯示卡記憶體,每顆晶片 32 條資料線,模型裡畫成一個車道。
Model drawn by My Culture Connect as a teaching diagram. The figures for each generation follow the table in Wikipedia, “High Bandwidth Memory”, which cites the JEDEC standards. · 模型由人師教育協會自繪示意;各代的數字依維基百科「High Bandwidth Memory」條目的規格表(引自 JEDEC 標準)。
Reading · 英文閱讀
Imagine the best cook in the world working in a kitchen where the refrigerator is at the other end of the street. The cook is fast, but most of the time is spent walking to fetch the eggs. A chip that runs AI has the same trouble. It can do an enormous number of calculations each second, but only if the numbers it needs arrive in time.
Those numbers are kept in memory chips. In an ordinary computer, the memory sits in its own packages some distance from the logic chip, and each memory chip talks through only a few dozen data lines. For everyday work that is fine. For AI, which moves mountains of data, the logic chip ends up waiting. The memory cannot deliver fast enough.
HBM was made to fix this. The letters stand for High Bandwidth Memory. Bandwidth means how much data can move in one second. Engineers did three things. First, they stacked memory chips on top of one another, like the floors of a tower. Second, they ran tiny copper posts, called TSVs, straight through the stack, like elevators, so every floor is connected.
Third, they moved the tower right next to the logic chip, on the same interposer you met in the last lesson. Because the two are so close, they can be joined by a huge number of very short wires. The first HBM had 1,024 data lines for each stack. Think of a highway with more than a thousand lanes that is very, very short.
HBM became an industry standard in 2013, and it has grown quickly since. Each new generation sends data faster through every lane, and the towers get taller: 4 floors, then 8, 12, and 16. The newest standard, HBM4, doubles the highway to 2,048 lanes. One HBM4 stack can move about 2,048 gigabytes in a single second.
That is why you now hear about HBM in the news. An AI chip usually has several HBM towers standing around it, feeding it data all day. The memory is no longer far down the street. The refrigerator has moved right beside the cook, and it has a thousand doors.
Questions · 閱讀理解
Why does an AI chip sometimes have to wait?
What three things did engineers do to make HBM?
What does “bandwidth” mean?
Words & Phrases · 生字及片語
Quick Check · 小測驗
1. What do the letters HBM stand for?
2. What is bandwidth?
3. How are the memory chips in HBM arranged?
4. How many data lines did the first HBM stack have?
5. Why is HBM placed right next to the logic chip?
Bandwidth Calculator · 頻寬計算器
Suppose one movie file is 5 gigabytes. Choose a generation of HBM and how many stacks stand beside the chip, and see how many movies' worth of data they could move every second.
假設一部電影的檔案是 5 GB。選一代 HBM、選晶片旁邊放幾疊,看看它們每秒可以搬動相當於幾部電影的資料。
1 · Which HBM? · 哪一代 HBM
3 · Compare with a home internet line · 和家裡的網路比一比
The 5-gigabyte movie and the home internet speeds are examples. The gigabytes per second for each generation are the top figures in the standards for one stack; real products may run below them.5 GB 的電影和家用網路的速度都是舉例。各代每秒幾 GB 是標準裡一疊的最高數字,實際的產品可能比較低。
Movies moved every second · 每秒搬幾部電影
0 movies · 部
0 GB every second · 每秒這麼多 GB
Your home line would need — to send what this memory sends in one second.這些記憶體一秒送出的資料,家裡的網路要傳 —。
Try It in the Model · 在模型裡試試
Bandwidth is lanes times speed: the number of data lines multiplied by how fast each line sends data.
頻寬就是「車道數 × 車速」:資料線的數目,乘上每條線送資料的速度。
One lane, a long way. Watch how few cars reach the logic chip.
一個車道,路很長。看看到得了運算晶片的車有多少。
Four floors of memory, right beside the chip, with 1,024 data lines between them.
四層記憶體,就在晶片旁邊,中間有 1,024 條資料線。
Twelve floors, and each lane is more than six times as fast as in the first HBM.
十二層樓,每個車道的速度是第一代的六倍多。
The road doubles in width, and the tower reaches sixteen floors.
路寬了一倍,塔疊到十六層。
Go Further · 延伸閱讀
HBM by the Numbers · HBM 的數字
Six numbers from the HBM standards.
HBM 標準裡的六個數字。
HBM was adopted as an industry standard in October 2013.HBM 在 2013 年 10 月成為業界標準。
The width of the road for one stack in the first three generations.前三代,每一疊的「路寬」。
An ordinary graphics memory chip has 32 data lines. An HBM stack has 32 times as many.一顆一般的顯示卡記憶體有 32 條資料線,一疊 HBM 是它的 32 倍。
The tallest stack in the newest standards: sixteen memory chips, one on top of another.最新標準裡最高的一疊:十六顆記憶體晶片一顆疊一顆。
The 2025 standard doubles the road.2025 年的標準把路加寬一倍。
From 128 gigabytes a second in 2013 to 2,048 in 2025, for a single stack.單一疊的速度,從 2013 年的每秒 128 GB 到 2025 年的 2,048 GB。
HBM Stories · HBM 的故事
Four short stories behind three letters.
三個字母背後的四個小故事。
Bandwidth · 頻寬
Bandwidth tells you how much data can pass in one second. You can raise it in two ways: add more lanes, or make each lane faster. HBM does both. That is why it has “High Bandwidth” in its name.
頻寬告訴你一秒鐘能通過多少資料。要提高它有兩個辦法:增加車道,或讓每個車道更快。HBM 兩樣都做,所以名字裡才有「高頻寬」。
JEDEC · 2013
HBM was first developed by Samsung, AMD, and SK Hynix, and in October 2013 an industry group called JEDEC adopted it as a standard. A standard means that memory from different makers fits the same chips, the way any plug fits a socket made to the same standard.
HBM 最早是三星、超微(AMD)和 SK 海力士開發的;2013 年 10 月,叫做 JEDEC 的產業組織把它訂成標準。有了標準,不同公司做的記憶體就能配同樣的晶片,就像照同一個標準做的插頭都插得進插座。
4 → 16
The first HBM stacks had four memory chips. Later generations went to eight, then twelve, then sixteen. As the tower grows, each chip has to be thinner, and the tiny posts that run through them must line up from floor to floor.
第一代 HBM 疊四顆記憶體晶片,後來變成八顆、十二顆、十六顆。塔越高,每一顆晶片就要越薄,穿過它們的細柱也要一層一層對得剛剛好。
2025
In 2025 the world's largest makers of HBM were SK Hynix and Samsung of South Korea and Micron of the United States. The towers are then joined to logic chips by advanced packaging, the subject of the last lesson.
2025 年全世界最大的 HBM 製造商,是南韓的 SK 海力士和三星,以及美國的美光。這些塔接著要靠先進封裝(上一課的主題)和運算晶片接在一起。
Myth vs. Fact · 常見迷思
✗ Myth · 迷思HBM is a kind of AI chip that does the thinking.HBM 是一種負責思考的 AI 晶片。
✓ Fact · 事實HBM is memory. It holds the data and hands it over quickly. The calculating is done by the logic chip beside it.HBM 是記憶體,負責放資料、並且很快地交出去。計算是旁邊的運算晶片做的。
✗ Myth · 迷思High bandwidth just means the memory holds more.高頻寬就是記憶體裝得比較多。
✓ Fact · 事實Bandwidth is about speed of delivery: how much data moves each second. How much the memory holds is called capacity.頻寬講的是送得多快:每秒搬多少資料。能裝多少叫做容量。
✗ Myth · 迷思A faster logic chip always makes the whole computer faster.運算晶片越快,整台電腦就一定越快。
✓ Fact · 事實Not if it has to wait for data. A fast cook with a faraway refrigerator is still slow.如果它要等資料,就不一定。冰箱在很遠的地方,再快的廚師也快不起來。
✗ Myth · 迷思HBM is just ordinary memory with a new name.HBM 只是一般記憶體換個名字。
✓ Fact · 事實It is built differently: chips stacked with TSVs, set beside the logic chip, and joined by over a thousand data lines.它的做法不一樣:用矽穿孔把晶片疊起來、放在運算晶片旁邊、用上千條資料線相連。
Remember It · 記憶口訣
HBM is memory moved right next to the chip.
HBM 就是搬到晶片旁邊的記憶體。
Memory chips are stacked, and TSVs connect the floors.
記憶體晶片疊起來,矽穿孔連接每一層。
1,024 data lines in each stack, and 2,048 in HBM4.
每一疊 1,024 條資料線,HBM4 是 2,048 條。
That is what bandwidth means.
這就是頻寬的意思。
Safety First · 安全提醒
Classroom Activity 1 · 課堂活動一
You need · 準備材料
Steps · 步驟
HBM does both at once: far more lanes, and a much shorter road.
HBM 兩樣一起做:車道多很多,路也短很多。
Classroom Activity 2 · 課堂活動二
You need · 準備材料
Steps · 步驟
In a real stack, a great many posts must line up through every chip. That is one reason tall stacks are hard to make.
真的堆疊裡,非常多根細柱要穿過每一顆晶片、全部對準。這是高的堆疊很難做的原因之一。
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)。