ShadowEcho › Shadowing practice › English

This Could End The RAM Crisis

English shadowing practice · 180 sentences · 20:09 long · 2860 words

This Could End The RAM Crisis

About this practice session

This page turns the YouTube video “This Could End The RAM Crisis” into an English shadowing exercise. The recording is split into 180 short sentences that you can play one at a time, repeat after the speaker and record your own version for automatic pronunciation scoring.

The speaker talks at about 144 words per minute, a moderate pace, with an average of 15.9 words per sentence. That makes it suitable for intermediate learners who want natural, everyday speech.

How to practise with this video

  1. Play the first sentence and listen while reading the transcript. Notice the stress, the pauses and where the intonation rises or falls.
  2. Replay it and speak along with the speaker, copying the rhythm rather than reading word by word.
  3. Record yourself straight after the original. ShadowEcho scores each word for accuracy, fluency and completeness.
  4. Compare your recording with the original, retry the words marked in red, then move on to the next sentence.
  5. Practise 10–15 minutes a day. Coming back to the same video after a few days is an easy way to measure progress.

Transcript (English)

  1. This is a graveyard of semiconductor giants. And  right now, on opposite sides of the Yellow Sea,
  2. two new empires are rising from the ground.  One belongs to SK Heinix, the company building
  3. the world's most advanced AI memory. The other  belongs to CXMT, a Chinese company that until
  4. recently almost nobody heard of. Today, it's  building memory factories faster than anyone
  5. else on the planet. I've spent more than a decade  designing microchips. And when I first looked into
  6. that race, I thought every company is chasing the  same future, but they were not. One may become the
  7. next semiconductor empire. The other may join the  graveyard. Subscribe to the channel and let's find
  8. out. It is quite remarkable. The semiconductor  industry is full of success stories. Yet,
  9. the DRAM industry is full of bankrupt companies.  Unlike CPU or GPU, memory is largely a commodity.
  10. So customers don't care who manufactures it.  What matters is cost. As a result, the industry
  11. has become trapped in a cycle of expansion and  collapse. Demand rises. Companies race to build
  12. new factories and then capacity flats the market  and prices crash. Companies go bankrupt and the
  13. game repeats again. Just like Japan who once  dominated the memory industry with Toshiba,
  14. NC Hitachi and Fujitsu, today most of them are  gone. Europe had its own memory dream. Two,
  15. a company called Qimonda, a spin-off from  Infineon. The technology was excellent, but the
  16. timing wasn't. One brutal downturn was enough, and  Kim was bankrupt. The company disappeared, but the
  17. technology didn't. Years later, it resurfaced in  an unexpected place in China. There, CXMT acquired
  18. thousands of Qimonda patterns. Among them was one  of the most important inventions in modern DRAM,
  19. the buried world line architecture. And that is a  crucial part of the story. In older DRAM designs,
  20. the transistors sat on the surface of the  silicon. The world line, a tiny wire running
  21. across the surface, controlled the transistor  by turning it on and off. In other words, the
  22. transistor acted as the gatekeeper for the memory  cell. If this capacitor is charged to one volt,
  23. it's a binary one. And if there is no charge,  it's a binary zero. Here, the depth of silicon
  24. is used to store more charge while taking very  few area. And for decades, memory became cheaper
  25. as engineers simply made everything smaller. Then  that strategy began to fail. But eventually there
  26. simply wasn't any room left to keep shrinking it.  So something had to move. So engineers moved word
  27. line underground below the surface of the silicon.  And suddenly more space became available above
  28. it. Essentially instead of building outward they  started building downward. And this allowed memory
  29. scaling to continue. But it came at a price.  Every new generation demanded larger factories,
  30. more expensive tools, and billions of dollars  in investment. At the same time, memory prices
  31. continued to rise and fall in violent cycles.  One company after another disappeared, and after
  32. decades, the DRM industry consolidated to just  three key players. Samsung, SKH, and Micron. The
  33. industry became so traumatized by this boom and  bust cycles that survivors learn just one rule.
  34. Don't build too much. For decades, nobody dared  to break it. And then they all did. AI arrived.
  35. And almost overnight, the greatest industry fear  changed from not producing too much memory to
  36. suddenly producing way too little. As billions  of dollars poured into new data centers, demand
  37. exploded and one component suddenly became more  important than almost any other memory. Just have
  38. a look at this. A modern smartphone contains about  12 GB of memory. A high-end PC might have 32. Now,
  39. in 2021, Nvidia's amper GPU carried 40 GB of high  bandwidth memory. Then, Hopper contained 80 GB.
  40. Blackwell increased that 292 and now Rubin pushes  that number to 288 GB per GPU and that's 9 times
  41. as much memory as a high-end PC on a single chip.  Now zoom out. Companies don't buy a single GPU.
  42. They buy hundreds of thousands. And that means  one AI cluster requires millions of memory chips.
  43. And every new generation of AI hardware multiplies  the amount the entire industry has to manufacture.
  44. And that's why AI boom makes everything more  expensive. And that's why overcoming it will
  45. be very difficult and maybe not even possible. But  high bandwidth memory HBM used in AI is especially
  46. hungry for silicon because it's not a single chip.  It's an entire skyscraper of memory. In order to
  47. build a single HBM chip, we need to stack 12 to  16 memory dice, each thinner than a human hair,
  48. without cracking a single one. And stacking is  exactly where things become extremely difficult.
  49. Even Samsung and Micron have struggled  to manufacture the latest generations
  50. of high bandwidth memory because as the stack  grows taller, engineers encounter new problems
  51. like cracked dice, heat, and falling yields. Most  people assumed that if you know how to manufacture
  52. chips, you know how to manufacture memory. But you  don't. Memory is not a technology problem. It's a
  53. manufacturing problem. And it's a long marathon.  Every new generation teaches engineers thousands
  54. of lessons and the company that learns fastest  wins. And HBM makes this challenge even worse.
  55. If a single layer fails, the entire stack is lost.  And this means building more memory requires more
  56. wafers, more tools, more engineers, more clean  rooms, and ultimately more factories. And that's
  57. when we have to take a closer look at these  two construction sites. one in South Korea,
  58. the other in China. Calling them factories doesn't  really do them justice. They are more like cities
  59. built for the sole purpose of manufacturing  memory. They have their own power infrastructure,
  60. water treatment plants, chemical supply networks,  and thousands of workers. They are among the
  61. largest manufacturing projects on Earth. What's  fascinating is that these both companies saw the
  62. same future, the same opportunity, but came  to two completely different conclusions. One
  63. decided to build the fastest memory in the world  while the other decided to build as much memory
  64. as possible. The question is which strategy the  future will reward. SKHenix had won the first
  65. phase in the AI race, but now it had to face  an entirely different challenge. manufacturing
  66. enough memory to satisfy the richest customers in  the world and that meant building more factories.
  67. This is M15X. Calling it a factory doesn't really  capture the scale. It's closer to an entire city
  68. devoted to a single purpose manufacturing HBM.  Under one roof, SKHix is constructing three
  69. enormous fabrication facilities. The clean room  alone covers nearly 100,000 square meters and
  70. the air inside is so pure that even a single dust  particle is a problem. Most people assume that the
  71. most expensive machines inside the semiconductor  factory are the EUV systems used to print circuits
  72. in an HBM factory. Packaging equipment rivals it.  HBM depends on three critical technologies. TSVs,
  73. vapor thinning and bonding. Among them, the most  expensive and complex machines are TC bonders.
  74. They stack one dye on top of another with  heat and pressure so precise they can fuse 16
  75. layers of silicon without cracking a single  one. It costs up to $200 million per tool.
  76. And a single high bandwidth memory factory needs  a hundred of them. Are you still wondering why
  77. memory gets so expensive? And this brings us  to the fundamental difference between logic and
  78. memory factories. In logic, engineers obsess over  transistor complexity. In memory, they obsess over
  79. manufacturing precision. And the mood here isn't  the design, it's the process. That's why building
  80. the factory is only the beginning. You still have  to polish the recipe and push yields as high as
  81. possible. And to make the economics work, you need  enormous production volume. Because in memory, you
  82. don't make fortunes on individual chips. You make  pennies again and again, billions of times. Well,
  83. M15X is only a tiny piece of a much larger  plan. In Yongen, the company is constructing
  84. four additional fabrication plants. In Chongju,  it's expanding its HBM packaging capacity. And
  85. in Indiana, it's investing nearly $4 billion  in an advanced packaging facility. Together,
  86. these projects represent one of the largest  manufacturing expansions in the history of the
  87. semiconductors. SK Hynix looked at the future and  decided that it belongs to HBM. But what happens
  88. if you make an opposite bet? We will deep dive  into CXMT factory in a moment. Between running a
  89. startup, filming videos, and talking to customers,  an enormous amount of information passes through
  90. my head every single day. Ideas, features, action  items, meeting notes, and eventually it becomes
  91. impossible to keep track of all of that. And  taking notes isn't really the hard part anymore.
  92. The hard part is maintaining them. So, I've  been using something called Second Brain Note
  93. by GenSpark. The system has two parts. The first  part is the hardware. It's basically a very thin
  94. recording device that's about the thickness of a  credit card. So, I usually keep it attached to my
  95. phone or you can keep it in a wallet. It weighs  almost nothing. So, after carrying it for 2 days,
  96. I forgot it was even there. It has international  security certifications and a light that flashes
  97. while recording and supports 112 languages. One  charge lasts 35 hours, so I can basically charge
  98. it once a week and never think about it. It's  designed to capture the conversations, ideas,
  99. and little details that normally sleep by the  end of the day. But the most interesting part
  100. is the software behind it. You connect Second  Brain to tools you already use such as Gmail,
  101. Calendar, Slack, Notion, Google Drive, GitHub,  and others. And instead of having information
  102. scattered across 10 different applications, you  create a single memory layer across everything
  103. you do. Then I can ask questions like, "What  did we discuss with this person at our last
  104. call?" And it searches across meetings, emails,  documents, and notes to find the answer. And it
  105. doesn't just retrieve the information, it can  act on it. It can handle the administrative
  106. work that consumes so much of my time, like  drafting emails, following up after meetings,
  107. or drafting context scattered across dozens of  tools. Most AI systems are good at answering
  108. questions, but this one creates a second brain  that remembers. And in the age of AI, I think
  109. that's incredibly valuable because the biggest  advantage you have is your unique knowledge,
  110. your experience, and the connections you make  between the ideas. If you would like to try it
  111. yourself, GenSpark has a device now available  at 10% off, only $179 instead of usual $199.
  112. Check it out now through the link below. What  makes this so interesting? Across the Yellow Sea,
  113. CXMT looked at exactly the same opportunity and  chose a completely different strategy. Instead
  114. of building the fastest and the most advanced  memory in the world, they decided to build as
  115. much memory as possible. And that's because the AI  boom wasn't creating just one shortage, but two.
  116. HBM and conventional DRAM share much of the same  manufacturing infrastructure, the same fabs, many
  117. of the same process steps, the same engineers,  and crucially much of the same wafer capacity,
  118. which means every wafer committed to HBM is a  wafer that cannot be used for DRAM for consumer
  119. products. And CXMT saw an opportunity. Instead of  concentrating its resources on the world's fastest
  120. and most complex memory to build, the Chinese  company made a very different bet scale. At first,
  121. that seems completely irrational. After all,  HBM is where all the profits are. So why
  122. would anyone choose an opposite strategy? The  answer lies in manufacturing. While SK Hanix,
  123. Samsung and Micron shifted more and more capacity  towards HBM, CXMT had doubled down on conventional
  124. memory. The chips that power laptops, smartphones,  servers, and data centers. And almost all of that
  125. expansion is happening in a single place, Hay.  Unlike SK Hynix who spent decades accumulating
  126. manufacturing experience, CXMT inherited some  patents from abroad including technology linked
  127. to Kim. But blueprints are one thing. Building a  worldclass DRM factory is another. Over the years,
  128. CXMT developed hundreds of its own patents. Yet  in memory industry, manufacturing excellence is
  129. everything. And it's built from millions of tiny  lessons accumulated over decades. And that's
  130. exactly what CXMT is set out to build in Hefe  from above. He doesn't look like the birthplace
  131. of a new semiconductor giant, but that's exactly  what it became. It began with a single facility
  132. known simply as FAB 1. Originally designed to  manufacture DDR4 memory, it now produces roughly
  133. 100,000 wafers every month. Then another factory  appeared. Then more and more new clean rooms
  134. were built. Piece by piece, an entire memory  manufacturing ecosystem emerged here. Today,
  135. the company produces roughly a quarter of a  million wafers every month. But that isn't the
  136. end of the story. CXMT is continuing to expand  the half a cluster with additional clean rooms
  137. and HBM development lines. By the end of 2028, the  production capacity is projected to approach half
  138. a million wafers every month. And that is truly  astonishing number because in less than a decade,
  139. China managed to build a manufacturer that  approaches the scale of Micron. That would
  140. place CXMT in roughly the same league as Micron,  a company that barely existed a decade ago. But
  141. this still leaves an obvious question. Why would  anyone choose an ordinary DRAM over HBM? Because
  142. HBM isn't simply a better DRAM. It's probably  the most difficult memory product ever put into
  143. mass production. Every new generation begins with  the most advanced DM nodes. Today, those are the
  144. one beta and one gamma generations. Producing them  increasingly depends on EUV lithography. SK Hynix,
  145. Samsung and Micron all use EUV to manufacture  their most advanced DRAM products. CXMT cannot, at
  146. least not yet. China has no access to EUV machines  and expert controls increasingly restrict access
  147. to many of the tools required to manufacture HBM.  And one consequence of this restrictions is that
  148. CXMT has been forced to pursue a different path.  Also, it has not abandoned HBM entirely. Quite the
  149. opposite. But CXMT also understands something  which SK Hynix learned decade ago that memory
  150. excellence cannot be inherited. It must be  learned. While SK Hynix was building upward,
  151. CXMT was building outward. So rather than  concentrating exclusively on HBM, CXMT is
  152. expanding the technologies it can manufacture  with the tools it has access to today. And that
  153. means mature DRAM process built largely around  DV lithography rather than the latest generations
  154. of EV systems like DDR5 and LPDDR5 produced in  enormous volumes because scale creates experience.
  155. experience improved yields yields lower costs and  those lessons eventually become the foundation
  156. upon which future generations of HBM can be  built. If we think about cultural differences,
  157. SKH and CXMT are building not just two different  factories, but they're also building it according
  158. to two completely different philosophies.  SK Hynix embodies decades of perfection and
  159. patience while CXMT represents speed, ambition,  and the determination to build an entire industry
  160. in a remarkably short period of time. Now you  understand why the AI race is no longer about
  161. algorithms and GPUs. It's becoming a race to  build memory factories and the market believes
  162. that story. Recently, both companies turned to the  capital markets. Skhinx raised billions of dollars
  163. to expand its HBM empire, while CXMT raised  about $8 billion in what became the largest
  164. semiconductor IPO in Asian history. So, who's bat  is smarter? At that point, I don't believe that
  165. this is the right question. I don't expect CXMT  to become the world's leading HBM manufacturer
  166. anytime soon, but I also don't think that's the  company's goal. CXMT isn't trying to beat SKH
  167. at its own game. It's trying to build something  much bigger and even more fundamental for China's
  168. domestic industry. Just think about it. China's  AI industry has continued advancing despite lack
  169. of compute. And just imagine what happens once  those constraints begin to disappear. And given
  170. the amount of capital, talent, and effort being  put into that development, this moment can come
  171. sooner than many people expect. And if that  happens, the entire semiconductor landscape
  172. could change. For a long time, the market has  been dominated by three companies. Suddenly,
  173. there could be four. and then millions of  additional wafers could flood the market which
  174. is wonderful for consumers but in this industry  excessive supply has a habit of destroying
  175. fortunes so everyone expected Chinese memory to be  cheaper instead the opposite just happened DDR5 is
  176. now more expensive than Samsung's and as we are at  half a million mark I just want to thank you for
  177. watching the channel for your amazing support  and as As a gest of appreciation, I want to
  178. give away some of these beautiful, beautiful  wafers. If you would like to get one, share
  179. this video with two of your friends and leave a  comment below why do you watch this channel and
  180. what videos you would like to see next. Love  you guys and I will see you in the next one.

The original video belongs to its creator: watch it on YouTube.

More English shadowing practice