When you introduce CPU, GPU, TPU, DPU, and QPU at a party, CPU says 'I'm versatile', GPU says 'I'm visual', TPU says 'I'm trending', DPU says 'I'm data-centric', and QPU? Well, it's quantum; it knows all the answers, but when you look at it, it just says 'Maybe...'

man, you are so good at explaining things and combining memes. prodigy.

Multicore CPUs aren't the reason why we can run multiple applications at once. Operating systems could do that long before multicore CPUs were a thing. The technology which allows that is called process scheduling. The OS is basically switching between the running applications giving each of them a fraction of time (many times per second) to execute whatever code they are currently running. Having multiple cores just allows the OS to handle multiple processes more efficiently.

I asked the GPT to explain that video to my dog, he said, Explaining this to a dog might be a bit challenging since dogs don't understand complex technical concepts. However, you can use simple actions and treats to represent these concepts: 1. **CPU (Central Processing Unit)**: Show your dog a treat (representing a task) and then give it to them. CPUs are like you, the treat giver, handling everyday tasks. 2. **GPU (Graphics Processing Unit)**: Take a ball (representing a fun task), throw it, and encourage your dog to fetch it. GPUs are like a friend who helps with fun and sporty tasks. 3. **TPU (Tensor Processing Unit)**: Place a puzzle in front of your dog and watch them solve it (or attempt to). TPUs are like a clever friend who's great at solving puzzles. 4. **DPU (Data Processing Unit)**: Show your dog their leash (representing a connection) and take them for a short walk. DPUs are like a friend who helps you connect with others and stay safe during walks. 5. **QPU (Quantum Processing Unit)**: Okay, this one's tricky! Pretend to be a magician and hide a treat in one hand. Then, show both hands and let your dog choose. Sometimes the treat appears, sometimes it doesn't (representing quantum unpredictability). Remember, dogs won't grasp these concepts fully, but it can be a fun and interactive way to spend time with your furry friend! 🐶🐾

The thing with Quantum Computers is that, basically, you can take a function, feed it literally every possible input in one go, retrieve every possible output simultaneously, and then sample from that statistically to tell you things about the function. It's incredibly niche, and super hard to actually build, and we're not even sure how many problems can actually be accelerated with it, but it's fun to work through the Math for.

Seems like we have a family reunion here: CPU, the brainy one, GPU, the artist, TPU, the specialized smarty-pants, DPU, the traffic controller, and QPU, the one who says he's working, but you're always unsure because he lives in multiple realities!

You didn't mention DSP - digital signal processor. Specialized to run a single program continuously in a loop {input -> process -> output} with a strict time limit dictated by the sampling rate. Used extensively for stuff like audio, radio comm and oscilloscopes.

BRO! 6:16 was the MOST comprehensive visualization of matrix multiplication I've seen and I've watched several videos and tried reading several books because I've been self-studying and I came to the conclusion that I need a tutor or something but wow bravo. That actually made me excited to learn it! If you're not doing math videos can you point me in the right direction? Thank you and God bless!

"highly trained wizards called software engineers" gotta be one of the most accurate sentences said in history

CPU: We need to talk. GPU: Already calculated what you're about to say. TPU: Already completed your words with predictions. DPU: Already sent the message. QPU: I was on the next conversation.

Are we all gonna ignore the guy playing League of Legends with a controller at 5:01? 💀

6:50 somehow, touhou manages to appear everywhere

This is how history should be taught.

An analogy I really liked for comparing cpu to gpu is trains vs cars, cars (cpu) are really fast for transporting one person as fast as possible, while trains (gpu) are faster than cars when transporting thousands of people as fast as possible, a cpu has really low latency for executing one instruction while gpus abuse simd operations to reduce the average latency over many parallel and similar instructions.

3:03 Just to expand on this, CPUs don't strictly *need* multiple cores to run programs at the same time. What really allows this to happen is context switching and I/O. Iff you record the states of all of the registers and memory of a program (i.e. the program context), and then restore the registers and memory to those states at a later time, the program will continue running like no time passed. Operating systems use this nifty feature to constantly switch out the currently running program, and they do this so many times per second that the user feels like the programs are rhunning smoothly in parallel. And they switch contexts either when a certain number of milliseconds passes, or when the current program thread does some I/O, since a thread waiting for I/O to complete does not need the CPU.

One downside of a QPU is that you need to stay aligned, you don't wanna know what happens when you're QPU misaligned

It's important to clarify on most architectures (especially CISC) one clock cycle usually isn't one instruction, only some very fast instructions can execute in one clock cycle but reading memory or division/multiplication can take several clock cycles.

QPU stands for Quadruple Parallel Universes

Its fascinating how graphics cards have come along Initially for graphics rendering Then crypto Now Deep learning neural networks and Ai Wow i wonder what they'll do next

Blown away by how good this content is. Thanks!

Correction: 1 Hz does not mean 1 instruction per second. Many types of instructions, like multiply/divide and memory operations, take multiple clock cycles to finish. 1 Hz just mean its clock runs once per second. Edit: I'm not completely sure about this second one, but I think Neumann in von Neumann is pronounced NOY-min, not NEW-min.

OMG you're the best at explaining tech topics in a digestible, memeable format

Your videos are epic! The information and humour density are perfect!

TPU is a brand-specific chip, made by Google for Tensor Flow. TPU is not a standardized term, but instead AI Accelerator would be used instead.

1:26 this is the ideal computing output, you may not like it but this is how peak performance looks like

ALUs (arithmetic logic units) and FPUs (floating point units) also used to be a thing but now days are almost always part of the CPU (and are plenty powerfull so there is no need to add external ones).

This was the best video you’ve made yet. Keep it up!

This is the most comprehencive CPU vs GPU vs TPU vs DPU vs QPU guide I have ever seen

RISC-V can, like ARM, be used for both high performance and low power consumption.

I liked the outout 😅 1:27

Wow! This is by far the best layman video I've ever experienced on CPU variants. Keep doing these whoever you are.

And here I thought QPU stood for Quad Processing Unit Yes, that's a thing, usually a GPU-like parallel processor. Raspberry Pis mentioned in the video have one as part of their VideoCore "GPU". E.g. VideoCore IV has 12 cores, each processes 4 vectors per instruction cycle, with each vector being 4x32bit - so a 16-way 32bit processor. Used for the OpenGL implementation, but can also be programmed manually, if you dare (assembly or some higher level languages though with lesser support). It actually has decent branching support for a GPU, as you can mask away results for each vector element, and can branch completely if all your vector elements are fine with that branch.

they call me a YVPU -- youtube video processing unit -- because of my crippling addiction to watching youtube videos

The highly trained wizards, called software engineers 😂

Modern x86 processors does run fairly similarly to RISC type of processor but it still does have a lot more instructions in case they are still being used (to not break compatibility). RISC V will also be an interesting instruction set architecture but it is mainly just in microcontrollers and Raspberry pie type devices and not for personal and data center usage yet. There are a lot of special processors made like for taking pictures/video on phones and encoder/decorder. I would think with the rise of various machine learning models more processors will be made to optimize for them (or use FPGAs).

I think it's important to note that GPUs are much better at floating point operations which are essential to calculate co-ordinates for 3D rendering while CPUs are mainly good at integer operations - that's one of the reasons they co-exist and can't replace each other. I know the video explained some of it, but I'm surprised it didn't touch on the float-integer subject. Also, how did nobody point out the literal madlad playing League of Legends with a gamepad at 5:01, lol?

Who else is old enough to remember math coprocessors? My first computer as a kid was a 486SX in the mid 90s and I remember reading in the manual that an optional math coprocessor could be installed. I always wondered how much faster that would have made my machine

Awesome video! Thank you so much for all the good and accurate facts that i sure will use in the next months :D

that short animation just taught me how multiplying matrixes works

You could also go over older deprecated units. The FPU (floating point unit) which is now included in most CPUs, the short lived physics cards which now have merged with graphics cards and the long lived sound cards, when CPUs were not powerful enough to do sound and other functions at the same time. As you can see most of these units died due to COUs becoming more powerful and taking over their jobs. That is because there is a fairly hard barrier for performance needed for things like sound unlike graphics where the quality rose with performance of these cards.

You forgot the PPU (Picture Processing Unit) The old 80s 8 bit proto gpu. You'd typically find one on retro games consoles

3:04 Jeff made it sound like multicore processors are necessary to run more than one application at once. I don’t think he meant for it to come out that way, but just to clarify for anyone who is still new to this stuff… Multitasking has been possible for decades through a process called time slicing, where even on a single core, the CPU can do bits of work For multiple processes, effectively simulating things happening simultaneously even though things are still happening sequentially. It’s like if you are preparing a meal and you have multiple things on the stovetop at once, you check on one of them, bounce over to the next one and go between them until the meal is ready

Protect this man at all costs! Thank you for this explanation

Love everything aspect of this video, you have a gift

Even modern x86 CPUs use RISC under the hood. In stead they simulate x86 instructions with a hardware compatibility layer because the x86 instructions set has become so ridiculously complicated that implementing it directly into the base layer silicon was becoming a serious problem both for performance and circuit complexity.

The reason you can do multiple things at once isn't because of multiple cores, we could do it back with only a single cpu. Your CPU does a tiny bit of computation at a time for multiple processes, and switches between them rapidly. Think of it like a chef cooking 20 meals at once. The reason this isn't noticeable is largely due to the vastly slower I/O commands it has to wait for. You can think of this as frying. You can think of the CPU as cracking the egg, plating up, etc.

This intro was so simple , fun and creative explanation of what is a computer

The videos just keep getting better!

Inside the Broadcom Videocore GPU there are 4 slices. Each slice contains a quad processing unit. Which is a quad threaded CPU risc CPU core with additional instructions for graphics operations. It runs a proprietary firmware based on Microsoft threads and is also responsible for the boot sequence on the raspberry pi.

the TPU is called the template processing unit. it is a chip specifically designed to speed up the compile times of c++ programs.

MAN ... i so rarely comment on things but your first 40 seconds were pure art in many forms ...

Man, i gotta say, i really like your way of explaining such complex subject with such a humor! The memes are hilarious

thank you, this video is fire!

That cpu outputting amouranth was the funniest thing I’ve seen all day.

Your videos are really informative , funny 😂😂 and short. Loved it.❤

That last "Trust me bro, it doesn't work" 🤣😂

4:47 Thats from an Nvidia graphic showing their older Pascal architecture on the left vs their newer Turing architecture on the right when it comes to matrix math

Pure Awesomeness. Thanks for posting this 😊

Another banger explainer video! Thank you 🙏

Modern GPUs actually have tensor cores included in them, so they're basically a GPU and TPU combined

Now there's NPU....

I just love the writing and explanations 💓

Bro this has to be the best opening out of any UKposts video I have ever seen

0:53 "built by Konrad Zuse in 1936 in his mom's basement" lol you're the best

1:26 I laughed so damnnn hard on this😂

I love how you make it sound in the beginning like it's some crazy task and all we do is write javascript.

Please make a full separate video on quantum computer. You are good in teaching complex concepts.

I think that I'm ready to get my degree after watching this video. My brain got literally overload with all that information at that pace.

You forgot NPUs. They will be used in Windows in the near future if you happen to have one inside your PC case. Taskmanager will also be able to recognize it.

You're the best man, love it!!!

That intro is fantastic! You just got a new subscriber! 😂

Great video

CPU to GPU: "You're pretty graphic, huh?" GPU to TPU: "You tensor to be dramatic, don't you?" TPU to DPU: "Always data-centered huh?" DPU to QPU: "Quantum of nonsensical bragging!" QPU: "I've processed this joke in a parallel universe where it's actually funny!"

This channel is the best thing I have discovered on UKposts.

I watched this on 2x speed but could fully comprehend due to having background knowledge but also the fact that your explanation is soo good! In 4 mins I understood this whole thing, thank you soo much !!

It would be nice ti have that output video complete and downloadable, for research reasons.

There is a common misconception that quantum computers will replace regular computers. If quantum computers ever become available to the general public (and assuming society doesn't collapse after our current forms of encryption get obliterated), then it is likely that we'd see QPUs working together with CPUs, in the same way CPUs and GPUs coexist - a QPU can do some tasks much faster (by using a different kind of algorithms that exploit quantum weirdness) but they would be much slower for other computations (using traditional algorithms like the ones CPUs and GPUs are optimized to run).

I wasn’t expecting a review of computer hardware history, but I’m here for it.

Ottimo video. Continua a pubblicare altre cose del genere.

Hey can you please make a video about the difference between ARM and x86?

I've gotten myself a Google Edge TPU USB stick, Coral Edge, which is super useful for some niche use cases. The power/energy efficiency makes it possible to let that run on battery too, interesting stuff

Fun fact - prior to parallel processing on multiple cores, only OS's that could handle time-slicing could run multiple programs at what, to us sloooooow humans looked like in parallel. The Amiga had such an operating system, and for a while was used by NASA to handle comms with satellites in preference to Windows boxen, because Windows at that time could literally only handle one program running at a time. On my A1200, I once had a word processor going, whilst OctaMED was playing some music, a little video cartoon (very low res by todays standards, mind!) was playing in another and a couple of other things that I can't recall what they were after all these years. That pushed my Amiga (an A1200 with a 68030 processor) to its' limits, and some of the processes would go a tad slow now and then, but OctaMED was chugging along quite nicely through it all. Sigh. I loved the Miggy, it was a joy to use. I so wish Commodore hadn't shot themselves in every limb before shooting themselves in the head!

Fireship videos are like gym for the brain, it feels good after watching it ❤

1:20 a cpu can do more than one instruction per 1hz it depends on transistor count

1:32 output 😂

This is a top quality material 🔥🔥🔥🔥

One of the best conclusions i've ever seen 👍

These videos are gold

6:05 Yes pls, I have a RTX 4080 and that is what I plan to do with it. Please provide me a video on how to train AI, ty. I'm not even fuckin' joking!

Man the thumbnail is Fire 😂 you are a Legend 👍🏼

Great vid, but I think you need one about FPGA's

All of that science and engineering so I can style a button with css.

This guy is just next level 😂 I love it ...im a new sub!

Jeff you’re the best teacher I’ve had this past decade

If you showed this video to someone 600 years ago they would start a new religion based off of it

One of the most interesting video ever!

That whole intro sequence was great

Man the starting was insane 😂, video is very helpful❤

0:16 electrical? you mean electronic?

There are also FPGA's (Field Programmable Gate Arrays) which are old and fascinating.. You set the instruction pipeline on that card beforehand, and feed it data. The output is a stream.