YouTube2Text

CPU vs GPU vs TPU vs NPU: What's the difference? — Transcript

by Bits & Bytes Unpacked · 2,147 words · 369 segments · language en · Watch on YouTube

Full transcript

  1. 0:00For decades, the CPU ruled computing.
  2. 0:03Every program, every operating system,
  3. 0:06every website, every spreadsheet, every
  4. 0:08painfully slow school computer lab, it
  5. 0:11all depended on one thing. The central
  6. 0:13processing unit. The CPU was the
  7. 0:15unquestioned king. Then, the world
  8. 0:18changed. Games became cinematic. Video
  9. 0:21became 4K. Cameras became computational
  10. 0:24photography machines. And artificial
  11. 0:26intelligence showed up with an appetite
  12. 0:28so absurd that it started consuming
  13. 0:30entire data centers for breakfast.
  14. 0:32Suddenly, the CPU was no longer enough.
  15. 0:35Now, Nvidia GPUs are treated like
  16. 0:37strategic resources. Google is building
  17. 0:40private AI chips in data center
  18. 0:41fortresses. Apple, Qualcomm, Intel, AMD,
  19. 0:45and Microsoft are shoving neural
  20. 0:47processors into every phone and laptop
  21. 0:49they can sell. This is not an acronym
  22. 0:51problem. This is a silicon arms race.
  23. 0:54CPU versus GPU versus TPU versus NPU.
  24. 0:58Four processors. Four different jobs.
  25. 1:00Four different strategies for
  26. 1:01controlling the future of computing.
  27. 1:04Let's begin with the old king, the CPU.
  28. 1:07CPU stands for central processing unit.
  29. 1:10It is the general-purpose processor
  30. 1:11inside nearly every device you own. Your
  31. 1:14laptop has one. Your phone has one. Your
  32. 1:17PlayStation has one. Even your smart
  33. 1:19fridge probably has one because
  34. 1:20apparently refrigerators now need
  35. 1:22software updates.
  36. 1:23The CPU is the manager of the entire
  37. 1:25system. It runs the operating system,
  38. 1:28launches applications, handles user
  39. 1:30input, moves data around, and tells the
  40. 1:33rest of the hardware what to do.
  41. 1:35If your computer were a kingdom, the CPU
  42. 1:37would be the king, the prime minister,
  43. 1:39the logistics department, and the
  44. 1:41exhausted intern answering emails at
  45. 1:432:00 a.m. It has to deal with
  46. 1:45everything. CPUs are built for
  47. 1:47flexibility. They have a relatively
  48. 1:49small number of powerful cores. These
  49. 1:51cores are designed to process
  50. 1:52complicated instructions, switch rapidly
  51. 1:55between different jobs, and make
  52. 1:56decisions based on changing information.
  53. 1:59For example, when you open a web
  54. 2:00browser, the CPU is involved in loading
  55. 2:03the app, managing memory, communicating
  56. 2:05with the operating system, responding to
  57. 2:07your mouse and keyboard, and
  58. 2:09coordinating everything that appears on
  59. 2:11screen.
  60. 2:12This kind of work is called sequential
  61. 2:14processing. One task may depend on the
  62. 2:16result of the previous task.
  63. 2:18You cannot load a web page before
  64. 2:20opening the browser. You cannot open the
  65. 2:22browser before the operating system
  66. 2:24starts. And you cannot pretend you are
  67. 2:26productive until at least three tabs of
  68. 2:28research and one tab of nonsense are
  69. 2:30open.
  70. 2:31The CPU is excellent at this.
  71. 2:34Intel Core processors, AMD Ryzen
  72. 2:37processors, Apple M series chips, and
  73. 2:40the CPU sections inside Qualcomm
  74. 2:42Snapdragon phone chips all perform this
  75. 2:44kind of general-purpose work.
  76. 2:47But the CPU has a problem. It is
  77. 2:49brilliant at handling complicated tasks.
  78. 2:51It is terrible at being asked to do
  79. 2:52millions of tiny repetitive tasks at the
  80. 2:55same time. Imagine hiring a
  81. 2:57Michelin-star chef to chop 10,000
  82. 2:59onions.
  83. 3:00The chef can do it, technically,
  84. 3:02eventually, but it is an expensive,
  85. 3:04dramatic, and deeply inefficient way to
  86. 3:06solve the problem. And that inefficiency
  87. 3:08created the opportunity for the GPU. GPU
  88. 3:11stands for graphics processing unit.
  89. 3:14Originally, it existed for one job,
  90. 3:16making video games look good. And for
  91. 3:18years, that was basically its entire
  92. 3:20identity.
  93. 3:22Want better shadows? Buy a GPU. Want
  94. 3:24realistic reflections? Buy a GPU. Want
  95. 3:28your game to render individual hairs
  96. 3:29blowing in the wind while your character
  97. 3:31walks through rain?
  98. 3:32Congratulations, your computer now needs
  99. 3:35a small power plant. But GPUs were built
  100. 3:37differently from CPUs.
  101. 3:40Instead of a few large, flexible cores,
  102. 3:42a GPU contains a huge number of smaller
  103. 3:44processing units. Each individual unit
  104. 3:47is not particularly smart. It cannot
  105. 3:49manage an operating system. It cannot
  106. 3:51make complicated branching decisions. It
  107. 3:53probably cannot explain why your Wi-Fi
  108. 3:55stopped working.
  109. 3:56But give it a simple mathematical
  110. 3:58operation, then give that same operation
  111. 4:00to thousands of other units, and
  112. 4:02suddenly it becomes terrifyingly fast.
  113. 4:05The GPU is not a master chef. It is an
  114. 4:07army. Imagine 10,000 workers on an
  115. 4:10assembly line. They cannot design the
  116. 4:11factory. They cannot run the company.
  117. 4:13But if every worker gets one identical
  118. 4:16task, they can produce an absurd amount
  119. 4:17of work in almost no time. That is
  120. 4:20parallel processing, and that is exactly
  121. 4:22what graphics require. Your screen is
  122. 4:25made up of millions of pixels. For every
  123. 4:27frame in a game, the computer may need
  124. 4:29to calculate lighting, color, texture,
  125. 4:32shadows, reflections, movement, and
  126. 4:35depth for huge numbers of pixels.
  127. 4:38Then, it has to do that again.
  128. 4:41And again.
  129. 4:42And again at 60 frames per second, or
  130. 4:45120, or 240 if you are competitive
  131. 4:49enough that losing is caused by
  132. 4:51insufficient refresh rate. A CPU would
  133. 4:54struggle to calculate all of that
  134. 4:56sequentially. A GPU can divide the work
  135. 4:58across thousands of processing units and
  136. 5:00calculate it simultaneously. That is why
  137. 5:03modern graphics exist. Nvidia's GeForce
  138. 5:06RTX cards, AMD Radeon GPUs, and the
  139. 5:11custom graphics processors inside
  140. 5:13consoles such as the PlayStation 5 and
  141. 5:15Xbox Series X all rely on this idea.
  142. 5:19Then came the plot twist. Researchers
  143. 5:21realized that the same kind of parallel
  144. 5:23mathematics used to render graphics was
  145. 5:25also extremely useful for artificial
  146. 5:28intelligence.
  147. 5:29AI systems, especially neural networks,
  148. 5:32perform gigantic numbers of simple
  149. 5:34mathematical operations. They process
  150. 5:37huge collections of numbers, multiply
  151. 5:39them, add them, compare them, repeat the
  152. 5:42process, and eventually learn patterns
  153. 5:44from massive amounts of data.
  154. 5:46That is exactly the kind of repetitive,
  155. 5:48highly parallel work GPUs were born to
  156. 5:50handle.
  157. 5:52So, the gaming chip became the engine of
  158. 5:53the AI boom.
  159. 5:55Nvidia did not just sell graphics cards
  160. 5:57anymore. It became one of the companies
  161. 5:59holding the keys to modern artificial
  162. 6:00intelligence. Its data center GPUs
  163. 6:03became critical for training large AI
  164. 6:05models.
  165. 6:06Companies building AI systems needed
  166. 6:08vast clusters of GPU hardware to process
  167. 6:11unimaginable amounts of data.
  168. 6:13The GPU went from the thing gamers argue
  169. 6:16about online to the thing governments,
  170. 6:18cloud providers, startups, and
  171. 6:20trillion-dollar companies are
  172. 6:22desperately trying to acquire.
  173. 6:24This is why high-end AI GPUs became so
  174. 6:26important. They are not just fast chips.
  175. 6:29They are industrial equipment for the AI
  176. 6:31age. But, relying on GPUs creates a
  177. 6:34problem. If everyone needs the same
  178. 6:36hardware, the company making that
  179. 6:38hardware gains enormous power.
  180. 6:40Google saw that problem early. And
  181. 6:42instead of accepting the GPU empire
  182. 6:44forever, it built its own weapon, the
  183. 6:46TPU.
  184. 6:48TPU stands for Tensor Processing Unit.
  185. 6:51It is Google's custom-designed AI
  186. 6:52processor.
  187. 6:54And unlike a CPU or consumer GPU, a TPU
  188. 6:57is not something you buy for your home
  189. 6:59setup. You will not walk into a store
  190. 7:01and say, "One Google TPU, please."
  191. 7:05These chips live in giant cloud data
  192. 7:06centers behind layers of security,
  193. 7:09cooling systems, networking equipment,
  194. 7:11and electricity bills that could
  195. 7:13probably fund a small country.
  196. 7:15The TPU exists because Google looked at
  197. 7:17AI workloads and asked a dangerous
  198. 7:19question.
  199. 7:20Why pay for a general-purpose processor
  200. 7:23when we can build a machine that focuses
  201. 7:24almost entirely on AI math?
  202. 7:27That is what a TPU is. It is a
  203. 7:29specialist.
  204. 7:31If a CPU is a king and a GPU is an army,
  205. 7:34a TPU is an automated factory designed
  206. 7:36for one product. No graphics. No
  207. 7:39operating system management. No
  208. 7:41pretending to be a gaming card.
  209. 7:43Its job is to process the specific
  210. 7:45mathematical operations used in machine
  211. 7:47learning as efficiently as possible.
  212. 7:50One especially important operation is
  213. 7:51matrix multiplication. That sounds like
  214. 7:54the kind of phrase that makes you
  215. 7:55immediately remember failing algebra,
  216. 7:58but it is central to how neural networks
  217. 8:00process information.
  218. 8:02AI models turn words, images, sound, and
  219. 8:05other information into enormous groups
  220. 8:07of numbers. Those numbers are repeatedly
  221. 8:09transformed through layers of
  222. 8:11calculations.
  223. 8:12A TPU is designed to accelerate that
  224. 8:14process.
  225. 8:16It sacrifices flexibility to gain
  226. 8:18efficiency. That is the trade-off. A GPU
  227. 8:21can handle graphics, AI, scientific
  228. 8:24computing, and many other parallel
  229. 8:26workloads. A TPU is much more focused,
  230. 8:29but when the job is the exact kind of AI
  231. 8:32math it was built for, specialization
  232. 8:34can be extremely powerful. Google uses
  233. 8:36TPUs in its cloud infrastructure for
  234. 8:39machine learning workloads. When you use
  235. 8:41AI-powered Google services, your request
  236. 8:43can be processed in huge data centers
  237. 8:45containing specialized hardware. That
  238. 8:48may include TPU clusters built to handle
  239. 8:50enormous volumes of AI computation. And
  240. 8:53strategically, this matters. Google does
  241. 8:56not want its AI future entirely
  242. 8:58dependent on another company's hardware
  243. 9:00roadmap. Building custom chips gives it
  244. 9:02more control over cost, supply,
  245. 9:04performance, and the direction of its AI
  246. 9:06infrastructure.
  247. 9:08The TPU is not just a processor. It is
  248. 9:11an insurance policy against someone else
  249. 9:13owning the factory. But massive data
  250. 9:15center chips are not the entire story
  251. 9:17because the next AI battlefield is not
  252. 9:20only in the cloud, it is in your pocket.
  253. 9:23NPU stands for neural processing unit.
  254. 9:25It is the smallest processor in this
  255. 9:27war, but it may end up being everywhere.
  256. 9:29The NPU is designed to run artificial
  257. 9:31intelligence tasks directly on your
  258. 9:33phone, tablet, or laptop. This is called
  259. 9:36on-device AI, and it is becoming a very
  260. 9:39big deal.
  261. 9:40For years, many AI features depended on
  262. 9:42the cloud. You take a photo, send it to
  263. 9:44a server, wait for processing, then get
  264. 9:46the result back.
  265. 9:48You speak into your phone, your voice
  266. 9:49travels across the internet, and a
  267. 9:51remote data center figures out what you
  268. 9:52said. That works until the internet is
  269. 9:55slow or unavailable or expensive or you
  270. 9:59would prefer not to send your face,
  271. 10:00voice, private photos, and personal data
  272. 10:03to a distant server every time you want
  273. 10:05a feature to work.
  274. 10:06The MPU is designed to change that. It
  275. 10:08runs trained AI models locally using far
  276. 10:11less power than forcing the CPU or GPU
  277. 10:13to do the same task. This is important
  278. 10:16because phones and laptops have
  279. 10:17something data centers do not,
  280. 10:19batteries.
  281. 10:20And battery life is brutally
  282. 10:22unforgiving. You can build the smartest
  283. 10:24AI feature in the world, but if it
  284. 10:25drains a phone from 80% to 12% while
  285. 10:28removing background noise, nobody is
  286. 10:30going to be impressed. The MPU is
  287. 10:33optimized for efficiency. It usually is
  288. 10:35not designed to train giant AI models
  289. 10:38from scratch. That kind of job still
  290. 10:40belongs to massive GPU and TPU clusters.
  291. 10:43Instead, the NPU runs models that have
  292. 10:45already been trained. This is called
  293. 10:47inference. For example, a company can
  294. 10:49train a massive speech recognition model
  295. 10:51in a data center. Then your phone can
  296. 10:53use a smaller optimized version of that
  297. 10:55model to recognize speech locally. The
  298. 10:57result can be faster, more private, and
  299. 10:59less dependent on an internet
  300. 11:00connection.
  301. 11:02This is why NPUs are showing up
  302. 11:03everywhere. Apple calls its dedicated AI
  303. 11:06hardware
  304. 11:08Qualcomm builds AI acceleration into
  305. 11:10Snapdragon platforms. Intel, AMD, and
  306. 11:13Qualcomm all offer processors with NPU
  307. 11:16capabilities for modern laptops. And
  308. 11:18recent AI-focused laptop categories
  309. 11:20increasingly require dedicated NPU
  310. 11:23performance for local AI features. The
  311. 11:25branding is different. The strategy is
  312. 11:28the same. Put AI hardware inside every
  313. 11:30device before someone else does. NPUs
  314. 11:33can help power features like facial
  315. 11:35recognition, camera enhancements, voice
  316. 11:38isolation, background blur during video
  317. 11:40calls, image classification, live
  318. 11:43transcription, and other AI-assisted
  319. 11:45tools.
  320. 11:46When your phone makes a dark photo look
  321. 11:48usable, detects a person in a frame, or
  322. 11:50removes background noise from your
  323. 11:52voice, specialized neural hardware may
  324. 11:54be doing part of the work. The NPU is
  325. 11:56not loud. It does not need a giant
  326. 11:59cooling system. It does not have the
  327. 12:01cultural status of a gaming GPU, but it
  328. 12:04may become the most personal AI
  329. 12:05processor of all, because it is the one
  330. 12:07living inside the device you carry all
  331. 12:09day. So, CPU versus GPU versus TPU
  332. 12:13versus NPU.
  333. 12:15Who wins? Nobody. Because this is not a
  334. 12:18battle where one processor destroys the
  335. 12:19others.
  336. 12:20It is a hostile division of labor. The
  337. 12:23CPU still runs the kingdom. It manages
  338. 12:25the operating system, applications,
  339. 12:28instructions, and the endless
  340. 12:30complicated chaos of general computing.
  341. 12:33The GPU owns massive parallel workloads.
  342. 12:36It renders games, accelerates graphics,
  343. 12:38and has become central to training the
  344. 12:40world's largest AI models. The TPU is
  345. 12:43Google's cloud-scale AI specialist. It
  346. 12:45is a custom-built answer to the
  347. 12:47question, "What if we stop renting the
  348. 12:49future from GPU companies?" And the NPU
  349. 12:52is the quiet invasion. It is putting AI
  350. 12:54directly into phones and laptops, making
  351. 12:57everyday features faster, more private,
  352. 13:00and less dependent on the cloud. The
  353. 13:01future is not one super chip that does
  354. 13:03everything. The future is a coordinated
  355. 13:06army of specialized chips, each taking
  356. 13:08the work it is best at. Your CPU gives
  357. 13:11the orders. Your GPU does the heavy
  358. 13:13parallel labor. Your NPU handles local
  359. 13:16AI quietly in the background. And
  360. 13:18somewhere in a giant data center, TPUs
  361. 13:20and enterprise GPUs are burning through
  362. 13:22enough math to make your old calculator
  363. 13:24file for emotional support.
  364. 13:27Anyway, that's the chip war. Like,
  365. 13:29subscribe, and stay tuned for the next
  366. 13:31episode of tiny rectangles deciding the
  367. 13:33future of civilization. Until then, keep
  368. 13:35exploring, keep learning, and never let
  369. 13:38the acronyms intimidate you.

About this transcript

This page contains the full transcript of CPU vs GPU vs TPU vs NPU: What's the difference? by Bits & Bytes Unpacked, generated from the public captions YouTube serves with the video. The transcript has 2,147 words across 369 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.