Why We're Reaching the Theoretical Limit of Computer Power

Video written by Amy Muller

[zfox]

As an electical engineer, this was probably the best explanation of transistors aimed at non-engineers I have ever heared. Props to Amy!

[joshuasims]

I like how quantum tunnelling kind of looks like the behavior of a loosely coded hobby physics engine. Sure, there's a collider, but if you're moving fast enough you go right through.

[badabingbadaboom]

1:20 "allegedly got those actors to the moon"💀 bro's wild for that

[biz]

it blows my mind how people figured out how to make computers

[novozal]

Can we all appreciate just how much of a legend Amy is?

[christophermyers]

I have a Ph.D. in physics and study theoretical chemistry for a living. Amy, you did an amazing job teaching yourself a topic that most people even with the proper training have a hard time describing! Being able to learn new and complex information is an impressive and valuable skill to have 😊. Nice job!

[AlexSchendel]

As others have noted, since the introduction of FinFETs (22nm process node) (FET stands for Field Effect Transistor), the "nanometer" naming of process nodes has become detached from any physical measurement of the transistor. Originally, it referred to the gate length of the transistor. However, as quantum tunneling and other current leakage started causing issues, gate length stopped shrinking as quickly. FinFETs enabled greater control over the transistor by surrounding the channel on three sides with the gate (rather than one side with planar).
Furthermore, it is important that there are two critical technologies which will keep Moore's Law alive:
1. New transistor architectures currently being developed:
1a) First, there is GAAFET (Gate All Around FETs) which enables greater control by surrounding the channel on all sides with the gate. It also flips the stack of fins on its side, so now channels are stacking vertically rather than horizontally. This naturally increases density further.
2a) Then we have fork GAAFETs which seek to further increase density by removing the size of the gap between an n-well and p-well in a pFET and nFET (all transistors come in positive and negative pairs and this makes them sit next to each other when they would normally have to be far apart to not cause issues with overlapping doped silicon - the phosphorous and boron mentioned in the video).
3a) Lastly, we have CFETs (Complimentary FETs) which seek to increase density even further by taking to two separate pFET and nFET stacks and stacking one on top of the other! Naturally, that could double density once again! As such, transistor density scaling is still far from dead. It has been getting exponentially more difficult and more expensive to continue, but there's still a lot of momentum left.

2. Advanced packaging: Packaging is the process of taking a silicon die (the actual silicon with all the transistors) and putting it in a package. For example, a CPU will take one or more dies, bond them to a green PCB known as the substrate, and then cover it with what is known as an IHS (Integrated Heat Spreader) to help dissipate the heat generated by the CPU. Originally, packaging just meant finding some way to connect your one silicon die to the outside world. However, as the desire for more compute and more transistors never stops, we had to get a bit more creative in packaging. This is because if you just make the silicon die larger, you end up wasting more silicon because we're cutting rectangular dies out of a circular silicon wafer. Larger rectangles means more lost silicon along the edges. Larger rectangles also means that a single little defect in the wafer (there's always some) will affect a larger piece of silicon. Better to have one bad die out of 100 rather than out of 20. So the question became: "How do we provide more transistors and more compute without further increasing die sizes and thus decreasing yields?" The answer was to split everything into multiple dies! So we started creating all sorts of techniques to put multiple dies next to each other on the substrate and allow them to communicate with each other. We even started stacking dies on top of each other! All of this allows us to keep dies small with high yields while not sacrificing on performance and transistor counts. And so as long as we can continue with advanced packaging, Moore's Law is still far from dead! This is because Moore's Law does not state how large of an area the transistors have to be in. He simply said that the number of transistors in an *integrated circuit* will roughly double every two years. If we can economically increase the size of that integrated circuit, then we can still increase the number of transistors on the integrated circuit as we please and thus continue adhering to Moore's Law!

While Quantum Computers are a very interesting topic of research and worth a video on its own, most people agree that it won't replace classical computers, at least not for a long, long time. This is because virtually all quantum computers require ridiculously well-maintained environments (fractions of a degree above absolute zero and state-of-the-art magnetic shielding) in order to not become completely useless due to interference. Furthermore, in order to create properly reliable and stable quantum circuits, you need thousands of qubits which we are well over a decade out from. Furthermore, nothing can beat the price of silicon. A CPU with tens of billions of transistors can be had for a hundred bucks. But on the mature process nodes, you can have microcontrollers that cost pennies. As such, even when we get proper quantum circuits, they will likely remain cost prohibitive for all except some of the most well-funded institutions.

[benmolitor]

This was an impressively accurate explanation of how transistors work. My only note is that Moores law is already dead, in that it refers exclusively to transistors per area. For production, instead of implementing the increasingly small transistors that science can create, we've been improving ~30nm technology for like a decade now (GAAFETS and FinFETS, for example). Industry terms like "5nm technology" are actually just jargon that means "a technology a few generations after 40nm technology that happens to give a performance increase similar to what 5nm technology would look like if it was possible for production right now."

[lonestarr]

Gotta say: Amy proves to be a great addition to the channel.

[justaguycalledjosh]

With how small transistors currently are, there is already some level of quantum tunnelling occurring, but at levels that existing error correction codes can detect and correct it.

[Netbase]

"allegedly got those actors to the moon"

hahaha I lost it.

[efovex]

3:57 As a physics major, that's the most accurate depiction of a quantum physicist I've ever seen.

[soffeebeans]

I took years of physics in college and specifically learned about the physics of transistors as I was in computer engineering. Amy did a better job in like 30 seconds than any of my teachers. That's seriously impressive, especially considering it's not even that simplified, and doesn't lose any actually important information. Hats off to Amy!

[nickjc]

I know a thing or two about quantum mechinics, and I've taken courses on quantum computing, so I have the pleasure of announcing: F in the chat for quantum computers, at least commercial ones in your home. They can do some tasks orders of magnitude better than regular computers, and other tasks not at all. So you can solve certain very hard math problems on a quantum computer very well (which isnt as useless as it sounds, since many things can be turned into math problems), but would never be able to, say, play minecraft.

And similarly to how there are quantum limits on how small transistors can get, there are quantium limits on how big quantum computers can get. Currently, thats at around 100 qubits (the quantum bit, rather than being 1 or 0, it can be a fun combination of 1 and 0), and even then you need to dedicate a third of them to error correction of your computations.

TLDR; sadge

[dfdemt]

2:18 - “….Amy, a SOCIOLOGY major”. The way Sam said that had me laughing really hard. 😂

[MFMegaZeroX]

As someone with a CS PHD, I feel a need to correct some incorrect parts of the video (and briefly talk about what the future holds). Warning: long and rambly

1) Transistor gate widths are actually about 32nm, not 3 nm. When the tech industry says "3nm process, " those aren't referring to real measurements. They are marketing terms used to continue naming new transistors as if they were scaling at the same speed as they used to. You can read actual dimensionality stuff in the more technical releases (or Wikipedia/WikiChip).
2) That being said, there is still issues of quantum tunneling of course, which is a big part of why we are about to stop shrinking transistors, but we still have until the 2031 cadence release until that stops
3) Speaking of which, after the 2031 transistor (probably will be called something like "1.0nm" or "10 angstroms"), the industry will only be fitting in more transistors by introducing 3D integration, adding a 3D element to our CPUs (which are currently logically flat). The gate width will probably be something like 16nm
4) 3D integration has a major heat issue though, which will require major redesigns of how we design our computers, which is part of the reason we are starting "power cores" versus "efficiency cores" since the heat problem means we can many slow efficiency cores, but not all problems are paralizable. So we still need some "power cores" for the things that are not. That will probably only get us another decade or 2 before even 3D integration with specific designs for it before the heat issues are too problematic (and even that doesn't happen, not too longer we'll reach a height problem, like flash memory will start seeing in a decade or 2).
5) Ideally, by that time some cool alternative things will propel us forward like better transistors (eg: carbon nanotube transistors) or alternative methods of computation.
6) No, quantum computing isn't a replacement. Realistically quantum computing will largely see use via cloud service and be targeted at researchers or companies for specialized problems that quantum computing is good at but and has a large input size.

[TheTexas]

I love that Sam is giving definitely not a slave Amy more publicity

[etaoinshr]

We NEED Amy on a Jet Lag season soon. With all these skills she's developed researching random stuff for HAI she'd mop the floor with the competition

[HeisenbergFam]

5:40 "humans are innovation gluttons making computers better until we die from them" one of the most accurate sentences said in history

[marvindebot]

I'm beginning to wonder if Sam has a basement like Simon but instead of writers, he keeps the interns and researchers there.