CPU vs GPU vs TPU vs NPU: What's the difference? — Transcript
Full transcript
- 0:00For decades, the CPU ruled computing.
- 0:03Every program, every operating system,
- 0:06every website, every spreadsheet, every
- 0:08painfully slow school computer lab, it
- 0:11all depended on one thing. The central
- 0:13processing unit. The CPU was the
- 0:15unquestioned king. Then, the world
- 0:18changed. Games became cinematic. Video
- 0:21became 4K. Cameras became computational
- 0:24photography machines. And artificial
- 0:26intelligence showed up with an appetite
- 0:28so absurd that it started consuming
- 0:30entire data centers for breakfast.
- 0:32Suddenly, the CPU was no longer enough.
- 0:35Now, Nvidia GPUs are treated like
- 0:37strategic resources. Google is building
- 0:40private AI chips in data center
- 0:41fortresses. Apple, Qualcomm, Intel, AMD,
- 0:45and Microsoft are shoving neural
- 0:47processors into every phone and laptop
- 0:49they can sell. This is not an acronym
- 0:51problem. This is a silicon arms race.
- 0:54CPU versus GPU versus TPU versus NPU.
- 0:58Four processors. Four different jobs.
- 1:00Four different strategies for
- 1:01controlling the future of computing.
- 1:04Let's begin with the old king, the CPU.
- 1:07CPU stands for central processing unit.
- 1:10It is the general-purpose processor
- 1:11inside nearly every device you own. Your
- 1:14laptop has one. Your phone has one. Your
- 1:17PlayStation has one. Even your smart
- 1:19fridge probably has one because
- 1:20apparently refrigerators now need
- 1:22software updates.
- 1:23The CPU is the manager of the entire
- 1:25system. It runs the operating system,
- 1:28launches applications, handles user
- 1:30input, moves data around, and tells the
- 1:33rest of the hardware what to do.
- 1:35If your computer were a kingdom, the CPU
- 1:37would be the king, the prime minister,
- 1:39the logistics department, and the
- 1:41exhausted intern answering emails at
- 1:432:00 a.m. It has to deal with
- 1:45everything. CPUs are built for
- 1:47flexibility. They have a relatively
- 1:49small number of powerful cores. These
- 1:51cores are designed to process
- 1:52complicated instructions, switch rapidly
- 1:55between different jobs, and make
- 1:56decisions based on changing information.
- 1:59For example, when you open a web
- 2:00browser, the CPU is involved in loading
- 2:03the app, managing memory, communicating
- 2:05with the operating system, responding to
- 2:07your mouse and keyboard, and
- 2:09coordinating everything that appears on
- 2:11screen.
- 2:12This kind of work is called sequential
- 2:14processing. One task may depend on the
- 2:16result of the previous task.
- 2:18You cannot load a web page before
- 2:20opening the browser. You cannot open the
- 2:22browser before the operating system
- 2:24starts. And you cannot pretend you are
- 2:26productive until at least three tabs of
- 2:28research and one tab of nonsense are
- 2:30open.
- 2:31The CPU is excellent at this.
- 2:34Intel Core processors, AMD Ryzen
- 2:37processors, Apple M series chips, and
- 2:40the CPU sections inside Qualcomm
- 2:42Snapdragon phone chips all perform this
- 2:44kind of general-purpose work.
- 2:47But the CPU has a problem. It is
- 2:49brilliant at handling complicated tasks.
- 2:51It is terrible at being asked to do
- 2:52millions of tiny repetitive tasks at the
- 2:55same time. Imagine hiring a
- 2:57Michelin-star chef to chop 10,000
- 2:59onions.
- 3:00The chef can do it, technically,
- 3:02eventually, but it is an expensive,
- 3:04dramatic, and deeply inefficient way to
- 3:06solve the problem. And that inefficiency
- 3:08created the opportunity for the GPU. GPU
- 3:11stands for graphics processing unit.
- 3:14Originally, it existed for one job,
- 3:16making video games look good. And for
- 3:18years, that was basically its entire
- 3:20identity.
- 3:22Want better shadows? Buy a GPU. Want
- 3:24realistic reflections? Buy a GPU. Want
- 3:28your game to render individual hairs
- 3:29blowing in the wind while your character
- 3:31walks through rain?
- 3:32Congratulations, your computer now needs
- 3:35a small power plant. But GPUs were built
- 3:37differently from CPUs.
- 3:40Instead of a few large, flexible cores,
- 3:42a GPU contains a huge number of smaller
- 3:44processing units. Each individual unit
- 3:47is not particularly smart. It cannot
- 3:49manage an operating system. It cannot
- 3:51make complicated branching decisions. It
- 3:53probably cannot explain why your Wi-Fi
- 3:55stopped working.
- 3:56But give it a simple mathematical
- 3:58operation, then give that same operation
- 4:00to thousands of other units, and
- 4:02suddenly it becomes terrifyingly fast.
- 4:05The GPU is not a master chef. It is an
- 4:07army. Imagine 10,000 workers on an
- 4:10assembly line. They cannot design the
- 4:11factory. They cannot run the company.
- 4:13But if every worker gets one identical
- 4:16task, they can produce an absurd amount
- 4:17of work in almost no time. That is
- 4:20parallel processing, and that is exactly
- 4:22what graphics require. Your screen is
- 4:25made up of millions of pixels. For every
- 4:27frame in a game, the computer may need
- 4:29to calculate lighting, color, texture,
- 4:32shadows, reflections, movement, and
- 4:35depth for huge numbers of pixels.
- 4:38Then, it has to do that again.
- 4:41And again.
- 4:42And again at 60 frames per second, or
- 4:45120, or 240 if you are competitive
- 4:49enough that losing is caused by
- 4:51insufficient refresh rate. A CPU would
- 4:54struggle to calculate all of that
- 4:56sequentially. A GPU can divide the work
- 4:58across thousands of processing units and
- 5:00calculate it simultaneously. That is why
- 5:03modern graphics exist. Nvidia's GeForce
- 5:06RTX cards, AMD Radeon GPUs, and the
- 5:11custom graphics processors inside
- 5:13consoles such as the PlayStation 5 and
- 5:15Xbox Series X all rely on this idea.
- 5:19Then came the plot twist. Researchers
- 5:21realized that the same kind of parallel
- 5:23mathematics used to render graphics was
- 5:25also extremely useful for artificial
- 5:28intelligence.
- 5:29AI systems, especially neural networks,
- 5:32perform gigantic numbers of simple
- 5:34mathematical operations. They process
- 5:37huge collections of numbers, multiply
- 5:39them, add them, compare them, repeat the
- 5:42process, and eventually learn patterns
- 5:44from massive amounts of data.
- 5:46That is exactly the kind of repetitive,
- 5:48highly parallel work GPUs were born to
- 5:50handle.
- 5:52So, the gaming chip became the engine of
- 5:53the AI boom.
- 5:55Nvidia did not just sell graphics cards
- 5:57anymore. It became one of the companies
- 5:59holding the keys to modern artificial
- 6:00intelligence. Its data center GPUs
- 6:03became critical for training large AI
- 6:05models.
- 6:06Companies building AI systems needed
- 6:08vast clusters of GPU hardware to process
- 6:11unimaginable amounts of data.
- 6:13The GPU went from the thing gamers argue
- 6:16about online to the thing governments,
- 6:18cloud providers, startups, and
- 6:20trillion-dollar companies are
- 6:22desperately trying to acquire.
- 6:24This is why high-end AI GPUs became so
- 6:26important. They are not just fast chips.
- 6:29They are industrial equipment for the AI
- 6:31age. But, relying on GPUs creates a
- 6:34problem. If everyone needs the same
- 6:36hardware, the company making that
- 6:38hardware gains enormous power.
- 6:40Google saw that problem early. And
- 6:42instead of accepting the GPU empire
- 6:44forever, it built its own weapon, the
- 6:46TPU.
- 6:48TPU stands for Tensor Processing Unit.
- 6:51It is Google's custom-designed AI
- 6:52processor.
- 6:54And unlike a CPU or consumer GPU, a TPU
- 6:57is not something you buy for your home
- 6:59setup. You will not walk into a store
- 7:01and say, "One Google TPU, please."
- 7:05These chips live in giant cloud data
- 7:06centers behind layers of security,
- 7:09cooling systems, networking equipment,
- 7:11and electricity bills that could
- 7:13probably fund a small country.
- 7:15The TPU exists because Google looked at
- 7:17AI workloads and asked a dangerous
- 7:19question.
- 7:20Why pay for a general-purpose processor
- 7:23when we can build a machine that focuses
- 7:24almost entirely on AI math?
- 7:27That is what a TPU is. It is a
- 7:29specialist.
- 7:31If a CPU is a king and a GPU is an army,
- 7:34a TPU is an automated factory designed
- 7:36for one product. No graphics. No
- 7:39operating system management. No
- 7:41pretending to be a gaming card.
- 7:43Its job is to process the specific
- 7:45mathematical operations used in machine
- 7:47learning as efficiently as possible.
- 7:50One especially important operation is
- 7:51matrix multiplication. That sounds like
- 7:54the kind of phrase that makes you
- 7:55immediately remember failing algebra,
- 7:58but it is central to how neural networks
- 8:00process information.
- 8:02AI models turn words, images, sound, and
- 8:05other information into enormous groups
- 8:07of numbers. Those numbers are repeatedly
- 8:09transformed through layers of
- 8:11calculations.
- 8:12A TPU is designed to accelerate that
- 8:14process.
- 8:16It sacrifices flexibility to gain
- 8:18efficiency. That is the trade-off. A GPU
- 8:21can handle graphics, AI, scientific
- 8:24computing, and many other parallel
- 8:26workloads. A TPU is much more focused,
- 8:29but when the job is the exact kind of AI
- 8:32math it was built for, specialization
- 8:34can be extremely powerful. Google uses
- 8:36TPUs in its cloud infrastructure for
- 8:39machine learning workloads. When you use
- 8:41AI-powered Google services, your request
- 8:43can be processed in huge data centers
- 8:45containing specialized hardware. That
- 8:48may include TPU clusters built to handle
- 8:50enormous volumes of AI computation. And
- 8:53strategically, this matters. Google does
- 8:56not want its AI future entirely
- 8:58dependent on another company's hardware
- 9:00roadmap. Building custom chips gives it
- 9:02more control over cost, supply,
- 9:04performance, and the direction of its AI
- 9:06infrastructure.
- 9:08The TPU is not just a processor. It is
- 9:11an insurance policy against someone else
- 9:13owning the factory. But massive data
- 9:15center chips are not the entire story
- 9:17because the next AI battlefield is not
- 9:20only in the cloud, it is in your pocket.
- 9:23NPU stands for neural processing unit.
- 9:25It is the smallest processor in this
- 9:27war, but it may end up being everywhere.
- 9:29The NPU is designed to run artificial
- 9:31intelligence tasks directly on your
- 9:33phone, tablet, or laptop. This is called
- 9:36on-device AI, and it is becoming a very
- 9:39big deal.
- 9:40For years, many AI features depended on
- 9:42the cloud. You take a photo, send it to
- 9:44a server, wait for processing, then get
- 9:46the result back.
- 9:48You speak into your phone, your voice
- 9:49travels across the internet, and a
- 9:51remote data center figures out what you
- 9:52said. That works until the internet is
- 9:55slow or unavailable or expensive or you
- 9:59would prefer not to send your face,
- 10:00voice, private photos, and personal data
- 10:03to a distant server every time you want
- 10:05a feature to work.
- 10:06The MPU is designed to change that. It
- 10:08runs trained AI models locally using far
- 10:11less power than forcing the CPU or GPU
- 10:13to do the same task. This is important
- 10:16because phones and laptops have
- 10:17something data centers do not,
- 10:19batteries.
- 10:20And battery life is brutally
- 10:22unforgiving. You can build the smartest
- 10:24AI feature in the world, but if it
- 10:25drains a phone from 80% to 12% while
- 10:28removing background noise, nobody is
- 10:30going to be impressed. The MPU is
- 10:33optimized for efficiency. It usually is
- 10:35not designed to train giant AI models
- 10:38from scratch. That kind of job still
- 10:40belongs to massive GPU and TPU clusters.
- 10:43Instead, the NPU runs models that have
- 10:45already been trained. This is called
- 10:47inference. For example, a company can
- 10:49train a massive speech recognition model
- 10:51in a data center. Then your phone can
- 10:53use a smaller optimized version of that
- 10:55model to recognize speech locally. The
- 10:57result can be faster, more private, and
- 10:59less dependent on an internet
- 11:00connection.
- 11:02This is why NPUs are showing up
- 11:03everywhere. Apple calls its dedicated AI
- 11:06hardware
- 11:08Qualcomm builds AI acceleration into
- 11:10Snapdragon platforms. Intel, AMD, and
- 11:13Qualcomm all offer processors with NPU
- 11:16capabilities for modern laptops. And
- 11:18recent AI-focused laptop categories
- 11:20increasingly require dedicated NPU
- 11:23performance for local AI features. The
- 11:25branding is different. The strategy is
- 11:28the same. Put AI hardware inside every
- 11:30device before someone else does. NPUs
- 11:33can help power features like facial
- 11:35recognition, camera enhancements, voice
- 11:38isolation, background blur during video
- 11:40calls, image classification, live
- 11:43transcription, and other AI-assisted
- 11:45tools.
- 11:46When your phone makes a dark photo look
- 11:48usable, detects a person in a frame, or
- 11:50removes background noise from your
- 11:52voice, specialized neural hardware may
- 11:54be doing part of the work. The NPU is
- 11:56not loud. It does not need a giant
- 11:59cooling system. It does not have the
- 12:01cultural status of a gaming GPU, but it
- 12:04may become the most personal AI
- 12:05processor of all, because it is the one
- 12:07living inside the device you carry all
- 12:09day. So, CPU versus GPU versus TPU
- 12:13versus NPU.
- 12:15Who wins? Nobody. Because this is not a
- 12:18battle where one processor destroys the
- 12:19others.
- 12:20It is a hostile division of labor. The
- 12:23CPU still runs the kingdom. It manages
- 12:25the operating system, applications,
- 12:28instructions, and the endless
- 12:30complicated chaos of general computing.
- 12:33The GPU owns massive parallel workloads.
- 12:36It renders games, accelerates graphics,
- 12:38and has become central to training the
- 12:40world's largest AI models. The TPU is
- 12:43Google's cloud-scale AI specialist. It
- 12:45is a custom-built answer to the
- 12:47question, "What if we stop renting the
- 12:49future from GPU companies?" And the NPU
- 12:52is the quiet invasion. It is putting AI
- 12:54directly into phones and laptops, making
- 12:57everyday features faster, more private,
- 13:00and less dependent on the cloud. The
- 13:01future is not one super chip that does
- 13:03everything. The future is a coordinated
- 13:06army of specialized chips, each taking
- 13:08the work it is best at. Your CPU gives
- 13:11the orders. Your GPU does the heavy
- 13:13parallel labor. Your NPU handles local
- 13:16AI quietly in the background. And
- 13:18somewhere in a giant data center, TPUs
- 13:20and enterprise GPUs are burning through
- 13:22enough math to make your old calculator
- 13:24file for emotional support.
- 13:27Anyway, that's the chip war. Like,
- 13:29subscribe, and stay tuned for the next
- 13:31episode of tiny rectangles deciding the
- 13:33future of civilization. Until then, keep
- 13:35exploring, keep learning, and never let
- 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.