Shadowing Practice: Every Processing Unit Explained in 20 Minutes - Learn English Speaking with Video

בחר מצב תרגול

כל מצב פותח חדר תרגול במסך מלא.

234 משפטים

אם המשפטים קצרים או ארוכים מדי, לחץ על ערוך כדי להתאים אותם.

CPU.

The CPU, or Central Processing Unit, is the main brain of every computer, phone, and console.

Every time you open an app, type a message, or click a button, the CPU is the one figuring out what needs to happen next.

What makes the CPU special is that it's a generalist.

It can handle almost any kind of task, from running your operating system to calculating the logic in your games.

A regular PC CPU usually has somewhere between 4 and 24 cores,

and each core is really good at solving complicated problems one step at a time, but that's also its weakness.

Think of a CPU like a few genius professors.

They can solve anything you give them, but if you hand them a million simple math problems at once,

they'll take forever, and that's exactly why every other processor on this list exists.

GPU The GPU, or graphics processing unit, was originally made for one job, drawing graphics.

NVIDIA popularized the name in 1999 with the GeForce 256, which it marketed as the world's first GPU.

Unlike the CPU, a GPU has thousands of small, simple cores.

Going back to our analogy, if the CPU is a few genius professors,

the GPU is a stadium full of students who can each only do basic math, but they all do it at the same time.

And that's perfect for graphics, because a 4K screen has about 8 .3 million pixels,

and the color of each one has to be calculated again for every single frame.

Oh yeah, that same power also turned out to be perfect for AI, because training AI is basically billions of simple math operations happening at once.

That's a big reason why NVIDIA became one of the most valuable companies in the world.

The downside is that GPUs are power -hungry and expensive, and they're bad at tasks that have to be done step by step.

PPU.

The PPU, or Physics Processing Unit, was a chip made only to calculate physics in video games, like explosions, flying debris, and ragdoll bodies.

It came out in 2006 from a company called AGEA,

as a separate card called PhysX that you plugged into your PC next to your graphics card.

The idea was cool, because physics calculations are heavy, and taking them off the CPU meant games could have way more destruction on screen.

But the card cost around $250 to $300 at launch, and only a handful of games supported it, like Ghost Recon Advanced Warfighter.

And in some of those games, turning it on actually made performance worse.

So in 2008, Nvidia bought Aegea and moved PhysX onto its own GPUs instead.

That's why you've probably never seen a PPU in anyone's PC.

The GPU basically swallowed it.

APU.

The APU, or Accelerated Processing Unit, is AMD's name for a chip that puts a CPU and a GPU together on the same piece of silicon.

AMD launched its first APUs in 2011.

The advantage is that you get decent graphics without buying a separate graphics card, which makes devices cheaper, smaller, and more power -efficient.

In fact, you've probably used one without knowing it,

because the PlayStation 5, the Xbox Series X and the Steam Deck all run on custom AMD APUs.

But since the CPU and GPU live on the same chip, they also have to share the same power, cooling, and memory.

So an APU is usually weaker than a separate CPU and graphics card.

By the way, Intel chips with built -in graphics work in a similar way, but only AMD calls them APUs.

SPU.

The SPU, or Synergistic Processing Unit, is a bit different from everything else here because it's not a standalone chip.

It's a type of core inside the cell processor, which was made by Sony, Toshiba, and IBM and used in the PlayStation 3 back in 2006.

The cell had one main core that worked like a normal CPU, plus eight SPUs that acted like a team of fast assistants doing heavy math.

One SPU was disabled to improve manufacturing yields, and one was reserved for the system so games could use six.

When developers used them properly, the PS3 could pull off things other consoles couldn't.

It was so good at math that the US Air Force built a supercomputer out of 1 ,760 PS3s.

The problem?

It was extremely hard to program.

A lot of developers struggled with it, which is why some multi -platform games looked worse on PS3 than on Xbox 360.

Sony dropped the cell for the PS4 and the SPU disappeared with it.

NPU.

The NPU, or Neural Processing Unit, is a small chip built specifically for AI tasks, and it lives right inside your phone or laptop.

Apple put one in the iPhone 8 and iPhone 10 back in 2017 and called it the Neural Engine.

You might think you don't use AI on your phone, but the NPU is working all the time.

It helps unlock your phone with your face, blurs your background in video calls, cleans up your photos, and powers features like live captions.

The reason it exists is efficiency.

The CPU or GPU could do these tasks too, but they would drain your battery and heat up your device much faster.

NPU power is measured in tops, or trillions of operations per second.

For example, Microsoft requires at least 40 tops for a laptop to get the Copilot Plus PC label.

However, NPUs are still too small to run the biggest AI models, like the ones behind ChatGPT, so those still run in the cloud.

TPU.

The TPU, or Tensor Processing Unit, is Google's own AI chip.

If the NPU is the small AI chip in your pocket, the TPU is the giant version that lives in Google's data centers.

Google started using TPUs internally in 2015, and in 2016 they powered AlphaGo, the AI that beat one of the world's best Go players.

A TPU is built around one specific type of math called matrix multiplication

which is basically the core math behind ai since it doesn't waste space on anything else it can often train

and run ai models more efficiently than a general purpose gpu

google trains its gemini ai on tpus

and in april 2026 it announced the eighth generation split into two chips the tpu 8t for training ai

and the tpu 8i for running it but you can't buy one of these and put it in your PC.

Companies rent them through Google Cloud, and some big AI companies like Anthropic have signed deals for huge amounts of TPU capacity.

LPU.

The LPU, or Language Processing Unit, is a chip designed by a company called Grok, and it's made for one thing, running AI chatbots really fast.

Grok was founded by Jonathan Ross, one of the engineers who helped create Google's TPU.

When a chatbot answers you, it writes the reply one small piece at a time, And the biggest slowdown isn't the math.

It's moving data from memory to the processor.

GPUs keep that data in separate memory chips, which is like a chef running to the fridge for every single ingredient.

The LPU keeps the data in super -fast memory called SRAM right on the chip, so the ingredients are already on the counter.

That's why Grok could generate text faster than many GPU setups.

The trade -off is that SRAM holds very little data, so big AI models have to be spread across hundreds of LPUs.

Still, the idea was good enough that in December 2025, NVIDIA paid about $20 billion to license GroK's technology and hire its leaders,

and in 2026 NVIDIA announced its own LPU racks designed to sit right next to its GPUs.

MCU The MCU, or microcontroller unit, is a tiny, cheap computer on a single chip.

It has its own small processor, memory, and storage all in one package,

which makes it perfect for simple devices that only need to do one job.

And they're literally everywhere.

Your microwave, washing machine, TV remote, car keys, and even your electric toothbrush are probably running on one.

A modern car can have dozens of them, controlling things like the windows, airbags, and brakes.

And if you've ever heard of Arduino, which is popular with students and hobbyists, That's basically a little board built around a microcontroller.

The downside is that MCUs are extremely weak compared to a CPU.

Many run at speeds measured in megahertz instead of gigahertz, and some have less memory than a single photo on your phone.

But that's the whole point.

They're cheap, they barely use any power, and some can run for years on a tiny battery.

ISP.

The ISP, or Image Signal Processor, is the chip that turns what your camera sensor captures into an actual photo.

And just to make this clear, a camera sensor doesn't take a nice picture on its own.

Each pixel on the sensor only sees one color, and the raw result looks dull, grainy, and unfinished.

So every time you press the shutter, the ISP jumps in and fixes all of that in a split second.

It fills in the missing colors, removes noise, adjusts brightness and white balance, and sharpens details.

100

In modern phones, it also helps combine several shots into one to create HDR and night mode photos.

101

This is why two phones with the exact same camera sensor can produce completely different photos.

102

A big part of the difference comes from the ISP and the software running alongside it.

103

Every smartphone has one, and so do digital cameras, webcams, security cameras, and even cars with driver -assist cameras.

104

Security processor A security processor is a small chip whose only job is to protect your most sensitive data, like passwords, fingerprints, and encryption keys.

105

On Windows PCs, it's called a TPM, or Trusted Platform Module, and on Apple devices, it's called the Secure Enclave,

106

which first showed up in the iPhone 5S in 2013.

107

Think of it as a vault with its own guard inside your device.

108

Your fingerprint or face data gets locked inside, and even the main operating system can't read it directly.

109

It can only ask the Volt simple questions like whether a fingerprint matches.

110

That way, even if malware infects your device, it still can't steal the secrets inside.

111

The TPM also checks that nobody tampered with your PC while it was starting up.

112

Oh, and this chip is also one of the main reasons a lot of people couldn't upgrade to Windows 11.

113

Microsoft requires TPM 2 .0, so many older PCs were left behind, even if they were still fast enough.

114

ASIC.

115

An ASIC, or application -specific integrated circuit, is a chip designed from scratch to do one single task and nothing else.

116

And technically, some chips on this list, like the TPU, are ASICs too.

117

The best -known example is Bitcoin mining.

118

At first, people mined

119

Bitcoin with CPUs, then with GPUs.

120

But starting in 2013, companies built ASIC miners designed only for Bitcoin's math,

121

and they were so much faster and more efficient that mining on a regular PC became basically pointless.

122

ASICs are also used in network equipment and in the chips that process videos on YouTube's servers.

123

But here's the catch.

124

An ASIC can't learn new tricks.

125

If the task changes, the chip becomes useless, and designing one on a modern manufacturing process can cost tens or even hundreds of millions of dollars,

126

so it only makes sense

127

when you need a huge number of chips doing the exact same job fpga an fpga

128

or field programmable gate array is a chip

129

that can be rewired after it's made a cpu runs software

130

but an fpga is a blank grid of logic blocks

131

that you can connect however you want

132

so it physically becomes the circuit you design think of it like a box of lego bricks

133

while every other chip is a finished toy the first commercial fpga came out in 1985 from a company called Xilinx,

134

which is now owned by AMD.

135

Because of this, engineers use FPGAs to test new chip designs before spending millions to manufacture the real thing.

136

They're also used in 5G cell towers and in high -speed stock trading, where saving a few microseconds means money.

137

And for gamers, FPGAs power retro devices like the Analog Pocket and the Mr. Project,

138

which recreate old consoles at the hardware level instead of emulating them.

139

The downside is that an FPGA is more expensive, slower, and less power efficient than an ASIC built for the same job,

140

and programming one is much harder than writing normal code.

141

DPU.

142

The DPU, or data processing unit, is a chip used in data centers to handle all the boring background work,

143

so the main CPU doesn't have to.

144

Servers in a data center are constantly moving huge amounts of data around, handling network traffic, managing storage, and checking security.

145

And all of that eats up a lot of cpu power

146

so the dpu takes over those chores think of the cpu

147

as a company's ceo without a dpu the ceo also has to sort the mail answer the phones

148

and check badges at the front door with a dpu there's a whole team of assistants for

149

that so the ceo can focus on the real work the most famous example is nvidia's bluefield

150

which came from nvidia buying the networking company melanox

151

and amazon uses a similar idea in its data centers with its own nitro chips

152

By the way, you'll never find a DPU in a home PC,

153

because it only makes sense when thousands of servers are talking to each other non -stop.

154

WSE The WSE, or wafer scale engine, is the biggest chip ever made, built by a company called Cerebrus.

155

And just in case you don't know, normal chips are made by printing lots of small chips onto a round silicon disk called a wafer,

156

and then cutting it into individual pieces.

157

Cerebrus doesn't cut it.

158

It turns the biggest square that fits on the wafer into one giant chip, and the result is huge.

159

The current WSE3 is about the size of a dinner plate, with 4 trillion transistors and 900 ,000 AI cores.

160

Because everything sits on one chip, data doesn't have to travel between separate GPUs through cables, which makes it extremely fast at running AI models.

161

That's why OpenAI signed a big deal with Cerebras in early 2026, and the company went public a few months later.

162

But building a chip this big is really hard.

163

Every wafer has tiny defects, so Cerebras adds spare cores to route around the broken ones.

164

It also needs its own special cooling and power system, so you can only use it as a full machine in a data center.

165

Neuromorphic Processor A neuromorphic processor is a chip designed to work like a brain.

166

A normal chip runs on a clock that ticks billions of times per second, and it keeps processing in fixed steps whether anything changed or not.

167

But your brain doesn't work like that.

168

Neurons stay quiet and only fire a signal, called a spike, when something actually happens.

169

Neuromorphic chips copy that idea with artificial neurons that only activate when needed, and they keep memory right next to where the computing happens,

170

just like the brain does.

171

You This makes them extremely power -efficient for things like processing sensor data in real time.

172

For example, IBM's TrueNorth chip from 2014 had one million artificial neurons,

173

and in 2024, Intel built a system called Holopoint using 1 ,152 of its Loihi 2 chips,

174

reaching 1 .15 billion neurons.

175

That's roughly the size of an owl's brain.

176

But neuromorphic chips are still mostly stuck in research labs they need completely different software,

177

and most of today's AI simply isn't built to run on them.

178

Photonic processor.

179

A photonic processor is a chip that uses light instead of electricity to do calculations.

180

Normal chips push electrons through tiny wires, which creates heat and wastes energy.

181

Light, on the other hand, travels incredibly fast, barely heats anything up,

182

and different colors of light can even carry separate data through the same path at the same time.

183

This is especially promising for AI, because light can do the kind of math AI relies on almost instantly just by passing through the chip.

184

In 2025, a company called LightMatter published a study in the journal

185

Nature showing a photonic processor running real AI models, including one that played classic Atari games like

186

Pac -Man, with accuracy close to regular chips.

187

However, light is great at moving and multiplying numbers, but bad at storing data, so photonic chips still need regular electronics next to them,

188

and converting between electricity and light costs energy.

189

That's why today, light is mostly used to connect chips together inside data centers, while fully photonic computers are still in the early stages.

190

QPU.

191

A QPU, or quantum processing unit, is the brain of a quantum computer.

192

Instead of normal bits, which are either zero or one,

193

it uses qubits, which can be in a mix of zero and one at the same time thanks to something called superposition.

194

Qubits can also be linked together through entanglement, so they behave like one connected system.

195

This lets quantum computers attack certain problems in ways normal computers simply can't.

196

In 2024, Google's willow chip, with 105 qubits, finished a test calculation in under five minutes

197

that would take one of today's fastest supercomputers an estimated 10 septillion years.

198

And in 2025, Google ran an algorithm on willow about 13 000 times faster than a supercomputer with results

199

that can actually be double checked

200

but here's the catch qubits are extremely fragile chips like google's

201

and ibm's have to be cooled to almost absolute zero colder than outer space

202

and even tiny disturbances cause errors also a quantum computer won't make your games

203

or your browser faster it's only useful for specific problems like simulating molecules for new medicines,

204

and both IBM and Google are aiming for truly reliable error -corrected quantum computers around 2029.

205

TSU.

206

The TSU, or Thermodynamic Sampling Unit, is one of the newest and strangest ideas on this list, developed by a startup called Extropic.

207

Normal chips spend a lot of energy fighting random electrical noise because noise causes errors.

208

The TSU does the opposite.

209

It uses that noise on purpose.

210

Instead of normal bits that are always exactly 0 or 1, the TSU uses probabilistic bits,

211

or pbits, which randomly flip between 0 and 1 because of natural heat inside the chip,

212

and you can control how likely each one is to land on either side.

213

That might sound useless, but a lot of modern AI, like image generators is built on probabilities and randomness,

214

so instead of a GPU burning energy to fake randomness, the TSU gets it almost for free from physics.

215

And unlike a quantum computer, it works at room temperature with normal transistors.

216

But keep in mind, Xtropic's big claims, like AI tasks running up to 10 ,000 times more efficiently,

217

come mostly from its own tests and simulations on small models.

218

Its first production -scale chip, the Z1, was only unveiled in 2026.

219

So for now, the TSU is a promising experiment, not a GPU replacement.

220

Biological Processing Unit And finally, the Biological Processing Unit, also called wetware, is a computer that uses real living human brain cells.

221

Scientists grow neurons from stem cells and place them on a chip covered in tiny electrodes,

222

which can send signals to the neurons and read how they respond.

223

Why brain cells?

224

Because the human brain is incredibly efficient.

225

It runs on roughly 20 watts, about as much as a dim light bulb.

226

While today's AI needs massive data centers, neurons can also learn and adapt on their own.

227

In 2022, a company called Cortical Labs taught neurons in a dish to play Pong,

228

and in 2026, its CL1 computer with around 200 ,000 human neurons learned to play

229

Doom, although it played like a total beginner.

230

The CL1 costs around $35 ,000, and a Swiss company called FinalSpark even lets researchers rent brain organoids online.

231

But of course, there are big problems.

232

The cells need constant life support with nutrients and the right temperature, and they only survive for a few months.

233

There are also serious ethical questions, like whether these cells could ever become aware of anything.

234

So for now, wetware is a research tool, not something you'll be putting in your PC anytime soon.

הורד אפליקציה

כל מה שאתה צריך כדי לדבר בשטף

הגיית AIציון לכל משפט

תרגול IPAשלוט בכל צליל

אוצר מיליםבנה את מאגר המילים שלך

משחק אוצר מיליםלמד תוך כדי משחק

סרוק QR

Download on theApp Store Get it onGoogle Play

FOR YOU

המלץ על סרטון

40 משפטים

אוצר מילים והערות הגייה לשיעור הזה

שיעור הדיבור הזה מבוסס על הסרטון “Every Processing Unit Explained in 20 Minutes”. המילים שחוזרות הכי הרבה: chip, cpu, gpu, processing, unit. בסרטון הזה יש 234 משפטים ו-3384 מילים לתרגול Shadowing. הדיבור נמשך 20:33. הדובר מדבר בקצב טבעי של כ-165 מילים בדקה, קרוב לשיחה יומיומית. רק 75% מהמילים נמצאות בין 3,000 המילים הנפוצות ביותר באנגלית, כך שאוצר המילים מאתגר.

אוצר מילים מרכזי בסרטון הזה

15 מילים מהסרטון שכדאי ללמוד, עם ההגייה והמשמעות שלהן:

מילה	הגייה	משמעות
chip שם עצם	/t͡ʃɪp/	שְׁבָב
quantum שם עצם	/ˈkwɑntəm/	קְוַנְט
efficient שם תואר	/ɪˈfɪʃənt/	יָעִיל
physics שם עצם	/ˈfɪz.ɪks/	פִֿיזִיקָה
server שם עצם	/ˈsɝvɚ/	שָׁרָת
calculation שם עצם	/ˌkælkjuˈleɪʃn̩/	חִשׁוּב \ חישוב
pocket שם עצם	/ˈpɑ.kɪt/	כִּיס
cloud שם עצם	/ˈklaʊ̯d/	עָנָן, עָב
solve פועל	/sɒlv/	פָּתַר
connect פועל	/kəˈnɛkt/	חיבר
genius שם עצם	/ˈd͡ʒin.jəs/	גָּאוֹן
artificial שם תואר	/ˌɑː.tɪˈfɪʃ.əl/	מְלָאכוּתִי
laptop שם עצם	/ˈlæp.tɑp/	מחשב נייד, מַחְשֵׁב נַיָּד
developer שם עצם	/dɪˈvɛləpɚ/	מְפַתֵּחַ
silicon שם עצם	/ˈsɪləˌkɑn/	צורן

Phrasal verbs שתשמעו

מילה	משמעות
figure out פועל	הֵבִין, מצא תשובה

דקדוק בסרטון הזה

המבנים שהדובר משתמש בהם הכי הרבה, עם המילים המדויקות מהסרטון:

מבנה	בסרטון
סביל (Passive voice) be + past participle — הדגש הוא על מה שקורה, לא על מי שעושה את הפעולה	was originally made · be calculated · be done
משפטי זיקה (Relative clauses) who / which + clause — מידע נוסף על אדם או דבר	students who can · card, which makes · processor, which was
הווה מושלם (Present perfect) have/has + past participle — פעולה שקרתה בעבר ועדיין יש לה חשיבות או השפעה עכשיו	you've probably never seen · you've probably used · have signed

הגייה ששווה לשים לב אליה

הדובר משתמש ב-28 קיצורים וצורות מקוצרות, כמו doesn't, can't, they're. אמרו אותם בצורה הקצרה, כפי שאתם שומעים אותם.

הצלילים של “th”: math /mæθ/, eighth /eɪtθ/, ethical /ˈɛθɪkəl/, algorithm /ˈælɡəɹɪðm̩/
הצלילים של “sh” ו-“zh”: efficient /ɪˈfɪʃənt/, calculation /ˌkælkjuˈleɪʃn̩/, artificial /ˌɑː.tɪˈfɪʃ.əl/, wash /wɑʃ/, efficiency /ɪˈfɪʃ.ən.si/
מילים ארוכות — שימו לב להטעמה: electricity /ɪˌlɛkˈtɹɪsɪti/, incredibly /ɪŋˈkɹɛdɪbli/, artificial /ˌɑː.tɪˈfɪʃ.əl/, developer /dɪˈvɛləpɚ/, complicated /ˈkɑm.plɪˌkeɪ.tɪd/

צלילים שדוברי עברית מתקשים בהם:

/w/ — שפתיים מעוגלות, לא /v/: quantum /ˈkwɑntəm/, willow /ˈwɪloʊ/, wash /wɑʃ/, wire /ˈwaɪɚ/, hardware /ˈhɑɹdˌwɛɹ/
/æ/ — פתוח יותר מה-“אֶ” העברית, לא כמו “a” ב-“cup”: math /mæθ/, graphics /ˈɡɹæfɪks/, calculation /ˌkælkjuˈleɪʃn̩/, lab /læːb/, laptop /ˈlæp.tɑp/
/r/ — הלשון לא נוגעת בשום מקום, לא ה-“ר” הגרונית: processor /ˈpɹɑˌsɛsɚ/, graphics /ˈɡɹæfɪks/, electricity /ɪˌlɛkˈtɹɪsɪti/, incredibly /ɪŋˈkɹɛdɪbli/, roughly /ˈɹʌf.li/

איך לתרגל עם הסרטון הזה

הקשיבו לכל הסרטון פעם אחת בלי לדבר, ורשמו את המילים שאתם לא מכירים.
התחילו במהירות 0.75×, תרגלו Shadowing משפט אחר משפט, וחזרו למהירות רגילה כשזה מרגיש קל.
הקליטו את עצמכם והשוו למקור, תוך תשומת לב למילים כמו chip, quantum, efficient.

מהי טכניקת השאדואינג?

שאדואינג היא טכניקת למידת שפה מבוססת מדע שפותחה במקור להכשרת מתורגמנים מקצועיים ופופולרית על ידי הפוליגלוט ד"ר אלכסנדר ארגואלס. השיטה פשוטה אך עוצמתית: אתה מקשיב לאודיו באנגלית ילידית וחוזר עליו בקול רם מיד — כמו צל העוקב אחרי הדובר עם עיכוב של 1–2 שניות בלבד. בניגוד להאזנה פסיבית או תרגילי דקדוק, שאדואינג מכריח את המוח והשרירים בפה לעבד ולשחזר דפוסי דיבור אמיתיים בו זמנית. מחקרים מראים שזה משפר משמעותית את דיוק ההגייה, האינטונציה, הקצב, הדיבור המחובר, הבנת השמיעה ושטף הדיבור — מה שהופך אותו לאחת השיטות היעילות ביותר להכנה לIELTS Speaking ולתקשורת באנגלית בעולם האמיתי.

טכניקת Shadowing: קרא את המדריך המלא שלב אחר שלב →

Shadowing Practice: Every Processing Unit Explained in 20 Minutes - Learn English Speaking with Video

בחר מצב תרגול

הורד אפליקציה

The Engineering that Runs the Digital World 🛠️⚙️💻 How do CPUs Work?

Why Memory Prices Are Crazy High

Mind-bending new programming language for GPUs just dropped...

John Carmack - The Elegant Design

The Only AI Tools You Need (12-Minute Guide)

Every Hardware Protocol Explained Simply in 10 Minutes!

Physics Informed Neural Networks (PINNs) [Physics Informed Machine Learning]

The Most Overengineered Part of Your iPhone

Did the AI bubble just pop?

Give me 31 minutes, I will make PyTorch click forever

Cache Systems Every Developer Should Know

Humans and multitasking - How much can we do simultaneously? | DW Documentary