Computer Timescales Mapped onto Human Timescales - Computerphile

[@ΝίκοςΙστοσελίδα]

Here is an idea for an analogy for branch prediction: You have to get out of your apartment and it might be raining. Your apartment is on the 10th floor and the only way to make sure is after going outside on the ground floor. Will you take an umbrella? If you do and it isn't raining, you have to pay a penalty by going back to the apartment to leave the umbrella there. If it is raining and you did take the umbrella, there is no need to go back to get it. Similarly, you can get examples for the rest of the cases.

[@justawatchin2]

As soon as Matt said he's in the US, I understood the hesitancy around the ballpark "15 minutes to the corner shop".

[@fburton8]

It's hard to get an appreciation of why computers are so _slow_ to e.g. open a project in Visual Studio when the hardware is so _fast_ .

[@paulmichaelfreedman8334]

Mostly security measures. An OS is 2 percent actual OS code, the rest is authentication and security. Bit exaggerated, but you get the gist.

[@fburton8]

@@paulmichaelfreedman8334 Does that apply even when the computer is off-network and the project source is purely local? I mean it _shouldn't_ , should it?

[@cromefire_]

Well Visual Studio is just slow because of M$... I mean have you ever zipped a folder in Windows Explorer? Use any 3rd party software and it's 10x faster, durch window completely and you probably get another 2x. I mean that's mostly filesystem latency, but I mean that fits well into the picture. If you aren't efficient with your file system and file access code you'll very often wait those 2.3 days it takes to go to the disk...

[@Cypeq]

there's an easy explanation, bad code.

[@Goob_V10]

It's because of bad programming. Plain and simple.

[@IkBenBenG]

I think the "1 frame of a game" example also quite well illustrates how much brute force a GPU has compared to a CPU. This CPU core would only be able to do 3 floating point add or multiply operations per pixel per frame. So if you want to multiply the R, G and B channels with some value, that by itself would already eat up almost your entire time budget if you want to target 1080p at 60Hz. But modern game engines often do hundreds or thousands of floating point operations for every pixel in every frame, and then still have to do other tasks like transforming and rasterising geometry as well. And on PC rendering at "only" 1080p and "only" 60Hz is nowadays often considered low-end. I know that GPUs achieve this trough mass parallelism instead of with raw speed, but I still find the magnitude of their raw compute power very impressive.

[@Themoonisachees]

i mean, given this specific task is massively parallel, the CPU could also do it faster by dropping precision and using SIMD instructions (especially since RGB values are integer), but yeah it's crazy to be writing a shader and simply ask the gpu "actually do that but in 1000 groups of 32 threads pretty please"

[@JohnBlackCyberSec]

Where are you getting the 3 floating point add operations per cpu core per frame from? That is most definitely not true. A single filled simd vector(xmm) on x86 can do 4 such additions in a single operations (for the smallest size simd vector). And a single core can execute these on multiple execution ports each of which can execute multiple of those per cycle....

[@VAXHeadroom]

@JohnBlackCyberSec I think that's for a filled pipeline...

[@hrayz]

@JohnBlackCyberSec If a CPU has to fill in each pixel of a 1920x1080 screen in 1/60th second. That's 1920x1080x60 = 124,416,000 pixels per second. So 1/x = 8.04 ns per pixel. Add or Multiplication (float), on the computer used in this video, took 2ns. So: 4 operations per Pixel is all the "time" the CPU is able to allocate.

[@Roxor128]

@@hrayz I just spent some time playing Deus Ex with the software renderer (so no GPU being used), and the Steam overlay was claiming 60FPS most of the time at 1920*1080 (the exceptions being around loading time). Admittedly, my machine is probably twice as fast as the example one.

[@JonBrase]

16:36 Using cache size numbers for Zen 5, L1 is roughly equivalent to a 28-sheet packet, 8.5"*11" paper size, printed front and back with 19 digit numbers in 10 point font with quarter inch margins. L2 is equivalent to a 289 sheet book L3 is 32 such books 64 GB of RAM is a library with roughly 130000 such books.

[@thekaxmax]

What about the international standard A4 page rather than Freedom units? And what font? The text size differs between fonts

[@mattilindstrom]

Half a nanosecond is about 15 centimeters of distance for light to travel in vacuum. For waveguides multiply that by 2/3 as a rule of thumb,.

[@kageofkonoha]

I remember from the show Reboot (animated show) in the 90's that the characters (who reside in a computer) always complained about how long 1 second is.

[@paulmichaelfreedman8334]

And todays computers consider one second a billion times slower than the dudes from that tv show.

[@viagra5207]

No more like ten to twenty times faster

[@robspiess]

Reminds me of the movie Tron from the early 80s measured distance in nanoseconds.

[@JDSileo]

Every time I have to restore a computer from scratch I imagine I'm doing what happens at the end of season 3 with the system crash... That show had such an influence on me that it made my career

[@farhanrejwan]

what people sometimes don't understand is just how much huge a thousand times of something is, compared to that thing. for example, 1 million seconds is about 11.57 days while 1 billion seconds is about 31.71 years. that's just how large 1 billion is compared to 1 million, not to mention just 1. and for computers, if 1 nanosecond was like how humans perceive 1 second in real life, then 1 second for them would be like how humans would perceive 31.71 years of time. that's how fast today's computers are, and it's going to be even faster in future.

[@chitlitlah]

I'm planning a big party for when I hit the age of 1.5... gigaseconds.

[@DavinaTheWeena]

Fantastic. When using my C64 I used to think about how quickly even that took to do things & that was mind boggling itself. E.g it could count to approx 860 in one second.

[@Turidus]

That made me curious, so I did a quick and dirty Java version on my machine. Completely unoptimised besides what ever the compiler provides. 20 run average: 3,68 * 10^8 counts per second. That's an increase by a factor of roughly 430000. Computers have become fast beasts.

[@r033cx]

It needs to be said that human brains are actually very fast, it's just that you as a person is an abstraction on top of your brain. It's like running an emulator on your computer, it's always way slower. It can also be compared to asking a neural network to add two numbers together. Even though its neurons do that internally all the time, it can't use them directly to do addition. In the same way you can't just use your fast neurons in your brain to crunch numbers, you have to think of numbers as a concept in your head and manipulate that concept, follow specific steps to produce the result. If you could reconfigure your brain to do arbitrary calculations, I'm sure it would also go down to pretty low times.

[@Alex-fh4my]

True!

[@vinesthemonkey]

I could listen to Matt Goldbolt talk about optimization all day

[@kiseitai2]

Wait… is this the compiler explorer guy? His work is impressive. I love doing small c++ compilation experiments in compiler explorer!

[@vinesthemonkey]

@@kiseitai2 yes. me too. it even shows you roughly which lines assemble to what

[@markrussell5587]

he has a podcast you know ... "Two's Compliment"

[@existenceisillusion6528]

This is what I used to tell my students. If a clock cycle was scaled up to a second, it would be 10^9/(3600*24*365) approx 31.71 years. So if you were the CPU, it would look like the user is taking about 30 years to type in '2 + 2'. 😂

[@ignacionr]

What an excellent idea of an analogy. Very useful too!

[@JonBrase]

6:42 64 bits is roughly equivalent to 19 digits, 32 bits is roughly equivalent to 9 digits, so we should probably be using the time for a human to crunch a 19 digit addition as our baseline, and doing long division at 9 and 19 digit lengths.

[@JonBrase]

Assuming 4 digits takes 5 seconds, 19 comes out to about 24.

[@beskamir5977]

@@JonBrase I don't think it'd be a simple linear increase. Human brains tend to batch items together when there aren't too many of them, so I think that would make a 4 digit calculation considerably easier than a 19 digit calculation. The closest computer parallel to that feature would be SIMD, but it's basically that feature that lets us 'count' up to 4-6 items without actually having to enumerate each item. We can just look at a handful of items and know the answer. Pretty sure that would also apply when calculating large numbers and simply trying to ignore parts of the larger numbers might not work as well as just having the number already spaced out in batches of 3-4 digits like the numbers for a card or phone.

[@hoarderofsecrets]

Yes and No. It takes the computer the same amount of clock cycles if its adding two 1 digit numbers or if its adding two 19 digit numbers. So perhaps a fair compromise would be to average how long it takes a human to do each?

[@georgihristov4415]

Every second a computer can compute lifetimes of what a person can compute. William Shanks calculated 707 digits of pi over his lifetime. The first random python algorithm that I found on stackoverflow to compute digits of pi can compute 24000 digits in one second.

[@JoseJimeniz]

Tracing route to main memory: 1 0.0014 ms L1 Cache [4 cycles, the time it takes to do fmul or fma] 2 0.0069 ms L2 Cache [20 cycles, the time it takes to do a sqrt] 3 0.0138 ms L3 Cache [40 cycles] 4 0.0345 ms Main Memory [100 cycles] Trace complete. Reminder: CPU performance hasn't been about computing speed for a decade; it's about cache size.

[@celivalg]

Also, in 0.5 ns, light travels about 15 cm. From the video, 0.5ns is a clock cycle at 2GHz.

[@PurpleYoshiEgg]

i always thought performance was about computing things as few times as possible because the enterprise systems i manage seem to redo computation so much that caching the results in a database speeds things up more often than not ten fold

[@turun_ambartanen]

do you mean 0.0014 µs?!? The way you have it currently written: 0.0014ms = 1.4µs = 1400ns. 1400ns for 4 cycles is the equivalent to 2.8MHz On the other hand 0.0014µs = 1.4ns. 1.4ns for 4 cycles is the equivalent to 2.8GHz

[@notthedroidsyourelookingfo4026]

A modern 3D V-cache CPU with its 144 MB of total cache can store about a hundred floppy disks in them.

[@hammerth1421]

3D V-cache is awesome. It's kinda like HBM for paupers, but I still love it.

[@gasdive]

Half a nanosecond. So about the time it takes light to travel diagonally across the face of the phone I'm watching this on.

[@hammerth1421]

Yep. One of the limiting factors for the clock speed of a CPU beyond the thermal and voltage limitations to keep the chip from violently cooking itself is how far a signal can travel within one clock cycle at that frequency.

[@goodlookinouthomie1757]

I assume it would take a desktop pc with identical system specs longer than a phone to do the same calculation because the signals have to travel further.

[@GoronCityOfficialBoneyard]

Captain Jean-Luc Picard : How long a time? Lieutenant Commander Data : 0.68 seconds sir. For an android, that is nearly an eternity.

[@michaelpound9891]

Brilliant! My suggestion for rendering your own game by hand is to render a scene near the end of the level that’s fading to black.

[@anti-gravityseal7162]

Before I even clicked, I knew this would be Matt Godbolt. Loved this episode of Two’s Complement!!

[@MattGodbolt]

I forgot we'd done it on Two's CP too! 😊

[@SkullCollectorD5]

That Winamp throwback!

[@mwoliver]

That poor llama!

[@stenis0]

Throwback!?

[@robspiess]

@@stenis0 WACUP for the win!

[@DenDodde]

@@robspiess Ugh.. foobar is the way to go.

[@VAXHeadroom]

The first mini computer I worked on was a Data General Eclipse. There was only 64KB of memory, but the time-sharing was such that every login got basically the whole memory, and they were time-shared / swapped out periodically. Our swap disk was several MB, but the fancy thing was it had one head for EVERY TRACK. So there was no head movement, the only latency was the disk rotation and on average you'd be in a pretty decent place to read the data you want within 1/2 a rotation.

[@liamwelsh5565]

L1: Your desk L2: Your room L3: Your house RAM: Have to make a trip to the store Disk: Have to order it online

[@RonJohn63]

We were taught this at the beginning of my second comp sci course: registers are faster than RAM, which is faster than hard disks, which are faster than screens, which are faster than line printers. Long, long time ago...

[@brookead]

I love the analogy and it reminds me of when I used to teach a course on estimation. The example we used was multiply two 4 digit numbers together. I'll leave aside all the obvious questions students could have asked like "are both numbers 1000" or similar which obviously short circuits things and we always assume two, non trivial, four digit numbers which are randomly selected. We also assume no calculators. The point of the exercise is to show people how flippantly we will estimate (dramatically underestimate usually) a task which seems simple. When you then make the students actually DO the calculation after gathering all their estimates, you show them just how wrong they are. I have never ever seen anyone who wasn't training for a maths olympiad that could multiple two non trivial four digit numbers together in 20 seconds. It's typically well over a minute. The analogy can be mapped back to this video. The computer does it without any form of external memory in 4 cycles. 2 nanoseconds. The vast majority of humans with NO external memory (so no pen and paper to write down interim steps) would likely never get the answer, let alone the right one! And when you introduce pen and paper that's analogous to at best level 3 cache because of the time it takes to write down the question and the working out. If you have a particularly obnoxious group of students you do the exercise in three steps, where the first step was adding two four digit numbers together, to lull them into being cocky, because the result of that is at most a five digit number and the result of multiplying two four digit figures together is more likely to be an eight digit number. Then you do the multiply part. And then you ask them how long it would take to multiply three, non trivial, four digit numbers together, just to see how many people are still cocky and immediately just double how long it took them to do the two four digit numbers (you'd be surprised how many do exactly that, although the person who casually said 1.4 days was quite shocked when I interrupted his session the following day at the appointed hour to demand my answer!) Once, exactly once, did someone ask me if the first two of the three four digit numbers we were about to multiply were the two we had already done. One person noticed that we dropped the word random from the last part of the challenge... :)

[@YandiBanyu]

Oh man, that interrupting session to demand answer was gold! I wish some of my teacher was fun like that

[@rezdm]

A different comparison. Imagine you have ceiling of 3 meters in the room. You switch on the lights. The time it takes light to travel from a light bulb on the ceiling to the floor (3m) is about 10 nanoseconds. A modern computer will sum two numbers within just a single cycle, so at 2GHz that is 0.5ns. So a computer can sum up 20 pair of numbers on a single core, while light from the ceiling reaches the floor.

[@almicc]

Even though a human would have an equivalent 5 years to color in a 1920x1080 grid one pixel at a time, it still being just a minute per pixel is nuts. You couldn't go to sleep or do anything else or you would fall far behind.

[@NFSHeld]

This is also very useful to illustrate what role a "bottleneck" in overall speed plays. Like you can get a really fast CPU that doesn't need 20 seconds to add two numbers, but rather only uses 10 seconds, but that doesn't matter if your RAM is so small that you need to constantly load data from the SSD every other "hour" and wait for 2 days for the SSD to respond. Likewise it doesn't help if you have a huge and ultra fast RAM but your CPU is cheap and slow so that it needs weeks to churn through all the data in your RAM.

[@RealCadde]

I was expecting a comparison between a computer answering a question like "What's the distance to the sun" where the computer would have to ask the internet for that answer, where a human will recall that answer pretty much instantly at 1 au. Or in light minutes, which is 8. So the computer would spend 100+ years in computer time to find the answer to that question, but a human can do it in seconds. And even if the human doesn't know, the human can most definitely find the answer at the local library in under an hour. One could argue that we store all that information in our "softdrive" (grey matter) and so for a better comparison it would be like the computer calling up the SSD to fetch the answer. But no, we don't go to the corner shop to answer that question. We have it cached in our L1 memory. The computer doesn't. In other words, our cache memory is MASSIVE and we can recall answers to all sorts of things by memory alone and even pull up entire songs, video clips, images and other useless bits of information in mere seconds. Finally, ask a computer to identify the dog in a picture and that computer spends hours in human timescale to do it. Humans will see dog, click dog and have the rest of the day off.

[@mfr2]

👍🏻 I was expecting something like that, too!

[@humanmerelybeing1966]

Alternatively, how long would it take the computer to know it doesn't know the distance between the Earth and the Sun in comparison to asking a human who doesn't know? If the human has never learned that information, their answer of "I don't know" would be instantaneous (excluding any time spent trying to intuit or calculate the answer). But what if that human *did* know that fact at some point in their life? In that case, their memory of having once had that memory could cause them to take more time to answer. And even then, if they decide they can't remember, "1 AU" might pop into their head hours or days later entirely at random.

[@JaapVersteegh]

The king of Compiler Explorer. I clicked!

[@TheoneandonlyRAH]

this was amazing!!!! great episode!!

[@jurajhezel942]

Beautiful video! I'm absolutely mind blown here... 🤯 At times yeah I ponder about those millions and millions of 'operations' that a CPU does. Even more so the bandwidth that SSDs NVMe drives that can move gigabytes of data in seconds... Think how long would a human copy a 100 page book... And 100 pages are tiny when speaking in disc space... Technology is great it's stunning and humbling and inspiring... 😊

[@busterdafydd3096]

27:54 I think it's important to remember in the 16ms for frame drawing, that's not how long it takes to draw the frame. Most of that time will be taken up by updating “state variables”, I would bet it really would take less than 1 ms. And if you ask a human to try to do the equivalent, you provide a complete description of the current scene (state variables updated) and have them draw from that description.

[@humanmerelybeing1966]

Having just watched the new DLSS 4 stuff, this makes me wonder if that's why "AI" frame generation is so fast compared to "real" frame rendering.

[@mtBone83]

It Really Whips the Llama's A..

[@Rugops42]

I watched this at ×42 speed to try competing with a computer. Didn't understand a word of it.

[@jeromethiel4323]

"Something, something, something, something, Llamas ass." ^-^

[@sushidub]

Love this analogy! Good show gents.

[@korysovec]

Just shows how good the 3D cache is, that you can have 100MB of super fast memory, be it a bit further away than your standard L3, but still. No wonder those CPUs are so desirable.

[@mathman0569]

division by intigers is weird on computers, sometimes you can reduce it to multiplication, which is a lot faster. And modulo if you do it by a power of two can be reduced to a bit mask, which is about as fast as one can get. though really should should explicitly be doing bit masks to really take advantage of their speed

[@michaelmaguire4147]

I love that computers actually can't do division (correctly) and just kinda fudge it a bit, but the inaccuracy is at like the 0.00000000001 scale, so we (almost) never notice.

[@tachrayonic2982]

This reminded me of Tron Legacy, where Flynn said that 1 millisecond was equivalent to about 8 hours on The Grid. Assuming 5 seconds/cycle (Grid time), that's 5.76 cycles per microsecond. The Grid is nearly 350 times slower than the modern hardware in your example.

[@Lion_McLionhead]

Hard drive access is comparable to what it really took to access a book in another continent, in the age of sail. Reading Bach's latest hit in the new world took quite a seek.

[@Panboy2k]

I've never really liked this comparison of computer to human speed based on how many sums one can actively do, compare a computer running a LLM , and an IMG to text model and a robot ballance and control model, and a physic's simulation all at the same time, to a human doing maths while walking and see which performs better

[@DantalionNl]

Matt talking about L3 being maybe a megabyte, A yes 2010 what a time that was. Sure is astonishing how modern CPUs have 100+ megabytes of L3 at roughly twice the bandwidth of DDR5 memory.

[@enilenis]

What's amazing is how much computation a human can do on 100W of energy, when most of it is used on life support and only a fraction on processing. To reach brain's levels of power efficiency seems impossible. We can make smarter and faster computers, but they will also absorb thousands of times more watts per process and are as efficient as cars made of pure lead.

[@thekaxmax]

For now

[@FlyByPC]

Whoa -- there are TWO of us still using Winamp? Neat!

[@PaulsPubAndBrew]

I feel like a library would have been a great example. When they were trying to come up with office information time delays. L1 - you are holding a book. Find me the first word on page 50 L2 - same exercise but the book is sitting on a counter right in front of you L3 - same exercise but now I don't want the first word, I want the 50th word. RAM - the same exercise but you are in the New York City public library and need to use the card catalog system to locate the book Hard drive - same exercise but now you are starting in Los Angeles and can only use public busing transportation to get to NYC

[@marsh_usa]

Credits for that analogy go to Andrew S. Tanenbaum "Computer Networks". In the same book: Never underestimate the bandwidth of a truck. It's been decades since I read this book and will never forget those two analogies Edit: Thinking about it for some more minutes, it might also have been Hennessey and Patterson "Computer Architecture"... anyways, both books are reference works

[@LordHonkInc]

The first time it really hit me was when I learned about CPU architecture and the access speeds of different storage options. Before then, all I knew was "hard drive slow, SSD fast, RAM fastest," but learning about how fast CPUs access caches and registers, it's frankly hard to comprehend. Your PC could calculate the hundredth fibonacci number in the time it takes to access that number if it were stored in RAM (yes, to be precise the CPU fetches more than one number from RAM at a time, but let's assume you're just interested in that one).

[@frtls]

Is SMT or hyper threading more or less time expensive?

[@zzstoner]

My analogy of using the multiplication of 2 cycles (1ns) to be equivalent to 10sec (human time)... It would take a human roughly 316 YEARS to complete what a computer did in 10 seconds.

[@thekaxmax]

And the human would make at least a few mistakes, unless v careful

[@JosefdeJoanelli]

A 2hr movie would take 14,400 years to render by hand a. Loving Vincent was 65000 frames of individually painted oil paintings, created by 125 artists and took about 6 years in total, which is an actual 750 person-years of work, beating a computer by a factor of ~20. Not bad!

[@holobolo1661]

18:53 please do an episode on that, sounds very intriguing.

[@hammerth1421]

I once worked out that just my CPU - not even my GPU that's optimized for such things - could do about 10^8 vector multiplications per second, assuming ideal conditions. No idea if that's actually correct, I had to do vector multiplications by hand instead of checking my work on how many of those my CPU could do.

[@pierreabbat6157]

For me, multiplying 32 by 27 is a lot faster than multiplying 37 by 23, not just because it's 3-smooth but also because it was my address when I was a kid. For a computer, they take the same time, unless the 32 is a constant, in which case the compiler can compile a shift instruction.

[@mranthonymills]

My favorite way to visualize such small time slices is that light takes about 1 ns to travel 1 foot.

[@denischen8196]

Adding and multiplying floating point numbers is like working with numbers in scientific notation. Reading from SSD is like driving to a library in another state, getting all the books you need, staying a night at the hotel, and driving all the way back home. Fetching a piece of data from a local server is like taking a voyage across the Atlantic Ocean to another continent and back. How long does it take for a computer to solve an algebra problem?

[@lachvogel]

Great Video!!! Somewhere you are a factor of 10 off?? 22:47 - 10 ms are equivalent to 3.2 years but 25:37 - 10 ms are equivalent to 16 weeks?

[@Computerphile]

yes I think we had an slight error in the video (either in my graphics or in what Matt said) - will try to pin a comment correcting once I've spoken to Matt -Sean

[@ThisWorldMakesMeSad]

I love your videos :D

[@JavierAlfonsoBellotadeFrutos]

We want those " that is another episode"

[@domminney]

Is it pronounced SATA or SATA?

[@nixxonnor]

Light travels only 15 cm (6 inches) in 1/2 nanosecond... Electricity travels a bit shorter

[@fgregerfeaxcwfeffece]

Well, software rendering of a bitmap is exactly like the manually coloring the squares. Everything else not so much. But there is this one practical example where it works like that.

[@gudenau]

I would love videos on cache, NUMA and the new branch stuff in the new Zen CPUs.

[@TehPwnerer]

Run the trace route command if you ever want to know how many servers in between

[@Christopher-y5f]

Thanks for the forecast! I need some advice: I have a SafePal wallet with USDT, and I have the seed phrase. (alarm fetch churn bridge exercise tape speak race clerk couch crater letter). How can I transfer them to Binance?

[@user-iu1xg6jv6e]

Winamp with default skin?

[@mfr2]

Boring!

[@Royaleah]

I would like to see the art in a few years.

[@finnegon16]

great great great video!!

[@dominicbout]

To flip the analogy around: How long would a computer take doing 50 years worth of a human doing basic math (Considering an average of the cycle count of arithmetic operations?) 🤔

[@mikeysd]

Well explained. Thank you for this!

[@CalculusDogbertus]

In Star Trek TNG, the Borg Queen offered to make Data human. He considered it for 0.68 seconds which, he said, is an eternity for an android.

[@nyalldavis]

we have to do math to work out the memory address of a number in ram that we want. Because the caching function is magic, it knew memory address we wanted before we did.

[@ReedCBowman]

I would say L1 cache would be better analogized as the post-it note on your monitor that you keep writing stuff you think you're going to need over and over on.

[@nigelgunn_W8IFF]

It doesn't take a computer nanoseconds. The user has to access the calculator app, type in the two numbers, read the result.... In practice it's little quicker that pencil and paper.

[@wintermute111]

I always like to translate it to how many centimetres light travels every clock cycle

[@BytebroUK]

If you haven't done so already, you could quite easily segue this into a talk about mutexes and condition variables, you know 🙂

[@mattiaruoja]

Yeah, inevitably computer science nowadays returns to the good old days when it all kicked off, the past catches up more quickly than computers can evolve, interesting times btw 😊

[@Jopie65]

How long does it take for an LLM to come up with a single token? 😅

[@EyesOfByes]

20:27 And this is why it's so sad Intel canned the 3DXpoint tech.

[@AzCcc]

@ 8:05 As a human, there's no way I can multiply 2 floating numbers within the same duration it takes me to add 2 floating numbers 😅

[@harrytsang1501]

You want IEEE754 compliance or just some confidence interval?

[@electronash]

Dreamcast controller spotted! Has Matt been playing DCA3, by any chance? lol

[@jereziah]

What tool was used to ascertain the avg number of cycles per instruction here?

[@accueil750]

so cooll i lvoe this videooooo!

[@Herby-1620]

You need to demonstrate the speed of light. A nanosecond is about 1 foot of wire. So one cycle (500 pico seconds) is about 6 inches of wire.

[@polypolyman]

“Knowing what it looks like and filling it in” is basically what they’re trying to do with AI frame generation

[@Cypeq]

why is he like a decade behind on cache sizes in cpus

[@MattGodbolt]

I was using the laptop I'm on as order-of-magnitude sizes, 128KB/1M LB L2/8MB L3 (and I bet I also misspoke too...there's a lot of number there hah)

[@Cypeq]

​@@MattGodbolt yeah I mean I'm being a little nitpicky, you were just making an example after all, to explain the difference.

[@domminney]

Nice Space Potatoes music clip, I won’t demonetise you 😂

[@yellingintothewind]

Caches have balooned in recent years. L3 caches on modern ryzen systems are 32+ mb. Intel is going a similar direction.

[@supernenechi]

I regularly think about how fast computers actually are! Usually that thought pops up when I am waiting for a computer to slowly do something... If they're so fast, then why are they so sloooow!? 😂

[@Cereal.interface]

well damn... classifications and hierarchical or parallel orders or subclasses. what fo they mean?

[@PopescuAlexandruCristian]

From the computer's perspective, if you run js, Java, or c#, you mostly waste time waiting for data.

[@googleaccountuser3116]

The brain doesn't calculate even though electrons fly around in humans at the same speed with the additional advantage of a body that is perfectly adapted to the physical world. Able to filter information rather than store it locally.

[@FlyWR]

While I appreciate the scale comparison for in-cpu computation and external memory access, the initial idea for comparing human calculation vs computer calculation is extremely misleading. Because, to be even, the computer had to read pen-written numbers from paper (and not fed through a scanner, find it itself using something like a motorized webcam) or correctly recognize it from the audio. Also the calculation made by the human should be completely memorized (what's 1000+1000?), because modern CPUs don't actually add the numbers as humans normally do, they just use lookup tables. There are humans who did that memorization. And if all that is added together... human will most likely be faster than any modern (2024-2025) computer, and more precise (mostly due to recognition errors). So yea, computers are fast, but only when dealing with carefully prepared material.

[@holobolo1661]

amazing video, thanks

[@Whisper555]

20:45 Hi mum! 👋🏻

[@JoeRobertshaw]

"Super-click video-gamer person."

[@thomasmaughan4798]

Computers do not have a (universal) time scale. EACH computer has a specific clock frequency and different CPU chips require different number of clock cycles to perform an operation.

[@AshishJacobSamMCS]

In the perspective of computers, we (humans) are essentially frozen

[@swine13]

I am glad you specified that you were talking about humans when you said "we", just now.

[@chriskellett3533]

In the perspective of humans (me) my windows 11 work computer is frozen... again.

[@raffriff42]

To a computer, waiting for a human to hit OK or Cancel must seem like waiting for the return of the Great Prophet Zarquon.

[@no_name4796]

Not true. We are simply specialized in logical thinking, which computers are completely unable of, while computers are literally just machines brute forcing calculations by doing the exact same think bazillions times x second

[@alansillitoNYC]

I was thinking why does Matt, a UK guy, have a Target bag in his living room. Then he said he's in the US. Maybe I need a faster clock or more L1 cache 😂