All language subtitles for [English (auto-generated)] Every Processing Unit Explained in 8 Minutes!

CPU,

the central processing unit, the one

chip everyone's heard of, and the one

almost nobody can actually explain

without saying it's the brain.

Here's the thing.

A CPU is a generalist. It handles

anything you throw at it. Opening a

browser, running Excel, decoding a

video, checking if your mouse moved. And

it does it sequentially, one task at a

time, but at insane speeds. Modern ones

run at 3 to 5 GHz. That's billions of

instructions per second. Intel kicked

this whole party off in 1971 with the

4004,

a chip with 2,300 transistors.

Your current CPU has tens of billions

packed into a fingernail of silicon.

Today, you're looking at 4 to 24 cores

in a consumer chip, like Apple's M

series, Intel's Core Ultra, AMD Ryzen.

All these CPU has enough cores to make

your life smooth. Each core is basically

a mini processor that can run its own

task. Think of a CPU as that one kid in

school who was solid at math, history,

English, and gym.

Never the best at any single thing, but

you want them on every group project.

The catch, CPU are terrible at doing the

same small task a million times in

parallel.

Ask one to render a 4K frame and it'll

cry, which is exactly why we invented

the next one, GPU. Graphics processing

unit, originally built for video games,

accidentally became the most important

chip of the AI era.

Where a CPU has 16 billion cores, a GPU

has thousands of simple ones. It's the

difference between hiring one Nobel

laureate and hiring a warehouse full of

interns.

If the job is doing the same small math

problem a million times, the interns win

every time.

And it turns out rendering pixels,

training neural networks, and mining

crypto all look like that exact same

problem, massive parallel matrix math.

Nvidia's RTX 5090 packs over 21,000 CUDA

cores.

AMD's Radeon lineup fights in the same

weight class. These things chew through

ray-traced games, train GPT class

models, and melt your power bill. An RTX

5090 pulls 575 W under load. That's a

microwave.

But here's the dirty secret of the AI

boom, ChatGPT, Midjourney, Stable

Diffusion. None of them run on CPU. They

run on racks of GPU. Nvidia's market cap

crossed $3 trillion because the world

quietly needed a million of their chips

all at once. TPU.

Tensor processing unit. Google's answer

to, what if we built a chip that only

knows how to do AI math and absolutely

nothing else. Released in 2016, the TPU

was Google's admission that even GPU

were overkill for what AI actually

needed.

Neural networks are basically endless

matrix multiplications. So Google built

silicon that only does that at ludicrous

speed. No graphics, no physics, no video

decode, just tensors.

Think of a TPU as a NASCAR car. It can

only turn left, but on an oval track, it

humiliates any Ferrari.

Google's latest TPU v5p powers Gemini,

the AI in your search results, and

almost every neural network inside

Google's empire. You can't buy one. You

rent time on it through Google Cloud.

The tradeoff is flexibility. A TPU is

useless for anything outside machine

learning. But for the thing it was built

for, it's faster per dollar and per watt

than almost anything on Earth. NPU.

Neural processing unit, the quiet chip

sitting inside your phone and your new

laptop doing AI without nuking your

battery or phoning home to a data

center.

Apple slid the first serious one into

the iPhone 10 in 2017, the neural

engine.

Qualcomm followed with Hexagon. Now

every Copilot Plus PC is required to

have an NPU capable of 40 plus tops, or

trillions of operations per second, just

to earn the label. What does it actually

do? It's the reason your phone blurs

backgrounds in real time, transcribes

your voice offline, and unlocks with

your face in 200 milliseconds.

It's also why your new laptop can run a

local language model without the fan

sounding like a truck engine.

An NPU is basically a TPU's pocket-sized

cousin, smaller, dumber, but sips power

like it's rationed. When your iPhone

edits a photo instantly, that's the NPU,

not the CPU, doing the quiet heavy

lifting. LPU. Language processing unit,

a chip so new and so specific that most

people in tech still haven't heard of

it. Grok. Not Elon's Grok, this is the

chip company with a Q.

Built the LPU specifically to run large

language models at speeds nobody thought

were possible. We're talking 500 plus

tokens per second on Llama 3, where a

standard GPU setup might do 50. It

doesn't train models, it just runs them,

faster than anything else alive right

now.

The whole architecture is built around

the fact that language models are

sequential, word by word, token by

token, and GPU, for all their brute

force, weren't built for that rhythm.

If you've ever used Grok's demo and

watched the response fill the screen

instantly, like it's cheating,

yeah, that's the LPU. IPU. Intelligence

processing unit, built by Graphcore, a

British chip company that tried to

outthink Nvidia, and mostly ended up as

a cautionary tale.

The idea was genuinely brilliant.

Instead of GPU that do parallel math in

straight lines, built a chip that thinks

in graphs.

A spiderweb of 1,472

independent processors all talking to

each other at once. Great for certain AI

workloads. Problem is, the world

standardized on Nvidia's CUDA software

before Graphcore could get traction.

The IPU is technically impressive and

commercially struggling.

A reminder that the best chip doesn't

always win. The chip with the best

software ecosystem does.

DPU.

Data processing unit, the bouncer of the

modern data center.

Nvidia's BlueField, AWS Nitro, Intel's

IPU.

They all do the same boring essential

job, handle the infrastructure grunt

work so the CPU and GPU don't have to.

Networking, encryption, storage traffic,

security checks. In a modern cloud

server, the DPU is quietly shuffling

terabytes per second between machines

while the GPU focuses on the actual AI

model. You'll never interact with one

directly, but if you've ever used AWS,

Azure, or Google Cloud, a DPU handled

your packet. QPU. Quantum processing

unit, and this one's going to break your

brain a little.

A regular chip uses bits, zero or one. A

QPU uses qubits, which can be zero or

both at the same time, thanks to a

property called superposition. Instead

of checking every possibility one by one

like a CPU, a QPU explores all of them

simultaneously and collapses into the

right answer at the end. It's less a

faster calculator, more like asking the

universe to do your homework. IBM's

Condor chip crossed 1,121

qubits. Google's Willow chip, released

late 2024, solved a problem in under 5

minutes that would take the fastest

supercomputer on Earth 10 septillion

years. That's a number older than the

universe many times over.

The catch, QPU needs to be cooled to

near absolute zero. -273°C,

colder than outer space. They're the

size of a chandelier. They're wildly

unstable, and they're genuinely useless

for running Chrome. But for

cryptography, drug discovery, and

simulating molecules, they're about to

rewrite what's possible in the next

decade. So, here's the takeaway.

For your everyday computer, the CPU. For

gaming and training AI, the GPU. For

cloud AI at scale, the TPU. For AI on

your phone and laptop, the NPU. For

blazing fast language model responses,

the LPU. For data center plumbing, the

DPU. And for problems we currently can't

solve at all, the QPU is the answer.

Hey, I made a video on some of the most

important data structures. Definitely

check it out, it might really help you.

And with that, I'll see you in the next

one.