This New Semiconductor Could Revolutionize Computing
Vložit
- čas přidán 8. 01. 2024
- Researchers at the Georgia Institute for Technology have found a new semiconductor that’s a really good candidate for making computers faster and smaller than ever. Amazingly enough, it works by combining graphene with silicon carbide, to give a material with a sensible band gap that still has a high thermal conductivity.
Correction to what I say at 02:54 --- That should have been voltage, not current.
Paper here: www.nature.com/articles/s4158...
🤓 Check out our new quiz app ➜ quizwithit.com/
💌 Support us on Donatebox ➜ donorbox.org/swtg
📝 Transcripts and written news on Substack ➜ sciencewtg.substack.com/
👉 Transcript with links to references on Patreon ➜ / sabine
📩 Free weekly science newsletter ➜ sabinehossenfelder.com/newsle...
👂 Audio only podcast ➜ open.spotify.com/show/0MkNfXl...
🔗 Join this channel to get access to perks ➜
/ @sabinehossenfelder
🖼️ On instagram ➜ / sciencewtg
#technews #sciencenews - Věda a technologie
I worked for ASML from 1998 to 2010 and saw wafers go from 6 in to 12. Line size from 1 micron to 22 nm. If we can figure the graphene thing out, it will be very exciting!😊
I remember 1998, Windows 98, monica lewinsky. Pentuim 4 processors, big fat pc monitors. The RnB song Too Close by Next.
the science is also just hype
As a Georgia Tech alumnus, I get to say this advance is the result of the efforts of MY tribe! (You aren't supposed to notice that the majority of the authors are at Tianjin University in China.)
Thank you for clarification re Tianjin...
don't get too hyped. Universities and tech groups have been finding new methods to speed up computers in the lab for 70 years. Just cos it works in the lab doesn't mean it will mass manufacturable and cheap enough to be market competitive. Only a small fraction of inovations in the lab make it to mass consumer market.
Graphene + silicon carbide sounds too fiddly to use to produce billions of transistors per chip. Ok they got one working in the lab. Took manufactureres 50 years to go from one silicon transistor to 134billion per chip.
@@plasmaastronautexcept now we have the manufacturing technology that we used to mass produce chips with billions of silicon transistors. If we can repurpose or reinovate some of those techniques and ideas, then it may not take 50 years before graphene replaces silicon.
Also, @SabineHossenfelder. There are some new ideas emerging about being able to bond semiconductors directly to a diamond backplane for RF amplifiers! Can you imagine how fast it will be able to take the heat away? You can really push the envelope! :D
Interesting, hadn't heard about this. Thanks for mentioning!
Not to mention maybe because diamond is a good insulator for electricity maybe it would help to control noise and drift
Very cool thought. I didn't even think of that. Those double bonds sure come in handy! Great job SimEon!@@SimEon-jt3sr
We've been using diamond for at least 10 years now at the macro level in avionics to control the heat. The ridiculous hardness and expense makes it extremely fiddly to work with. For instance, you cannot tune a coil in a filter even a tiny bit without first removing the diamond paste. 100% guaranteed to rip the pad right off your board.
We've also had diamond being used as detection medium in some particle physics detectors for awhile now
An interesting video, and you explained almost everything correctly. I would quibble with your explanation that in a semiconductor “if you apply enough current” you can get the device to conduct. I think a better way to say it is that if you apply a voltage, you can bend the energy bands in the semiconductor, which creates a condition where electrons can aggregate and move freely to create a current. When you remove the voltage, the material reverts to being close to an insulator and the device “turns off.” As always, I really enjoy and appreciate your work!
Thanks for going into details about this, I was wondering how it worked
She's corrected it in the notes
Sabine may have been thinking of earlier bipolar junction transistors, but modern digital electronics relies on unipolar voltage switched metal oxide semiconductors.
@@Liberty4EverYou and Sabine have both used phrasing associated with distorted descriptions of popular transistor types used in microchips . Neither is the defining characteristic of semiconductor materials, merely what was built from them .
Not everyone is an engineer or a scientist here
This channel keeps us up to date ❤
2:52 there's a mistake! You apply a voltage (aka gate voltage) NOT a current, in order to allow for conduction (between source and drain).
And my comment is on the same topic. It seems like a certified quantum theoretical physicist can always say whatever he/she wants.
@@cavesalamander6308 Gates require a current even if only femto amps. Otherwise DRAM chips would be non-volatile ! Where do you think the charge goes ?
3:02 Typo "Voltage" , not current. To have electrons transit the energy band gap, a sufficiently high "Voltage" is applied across the gap.
As a semi conductor enthusiast, I am excited for this video
I truly thought Moore's law dead myself. However, it never ceases to amaze me what the motivation of large amounts of money and fame will inspire either. Hats off to Georgia, I didn't know they had it in them.
It's Georgia Tech, not Georgia, and Republican politicians in Georgia work overtime to damage this world-class institution.
Yee haw buddy, we got dem darn transistors stackable no problem y’all.
Wherever there is a law, somebody tries to break it. 😂
Someone ask Florida Man .
The answer may not be practical, but , it should be a hell of a show .
Well, Moore's law is dead. A new technology requires new laws in development. It will be a decade for the prototype, 5 years till production, 2 years to become profitable, 1 year until everybody and his dog know, how to make it "more efficient". So i expect this to be available to consumers in 20 years. And from this point onward, a new version of "Quantum-Moore's Law" COULD be applied. But i think, that's like "go faster" stripes on a car. They don't really make the car go faster.
I recall that around 1990 during the first years of micro mechanics, lecturers talked of using the production methods then invented to lithographically etch cooling channels in microchips so cold gasses could cool the electronics . The entire concept was easy manufacturing of the tiny channels using existing mass production equipment then connecting outside pipework during packaging along with the external gold wires .
What became of the idea?
Also, early '90s Motorola was building chips vertically to increase surface area of gates to increase power levels, while conserving horizontal space.
They've been declaring Moore's Law dead since I was a kid, and I'm north of 50. Faster clocks, multipatterning, stacking, multithreading, etc, keep pushing performance. Incidentally, the idea of transistors on carbon substrates predates graphene, with diamond semiconductors proposed in the early 90s. While expensive, they worked better as they got hotter up to about 700 C, and would not fail until over 900 C. And yet, we're still using SiO2 :P
Not to mention diamond is only expensive due to a monopoly/cartel. If diamond became the new oil, I'd imagine some heads would start rolling (perhaps literally) to dissolve it and reduce the artificially high cost.
@@WhiteDragon103the diamond they'd use would almost certainly be lab grown, to ensure uniformity and impurity control.
Moore himself is now dead too .
@@johndododoe1411💀💀💀
@@Ornithopter470 >almost certainly
> i.e. there's a chance they won't
what unlikely circumstances do u foresee whereby chip manufacturers would require natural diamond imported from African mines?
Sabine, truly excellent! I especially liked your setup of the problem this new material is intended to solve.
@SabineHossenfelder, nice to see an area i studied in my phd make its way to your videos and closer to reality. I published part of my thesis 'Role of substrate induced electron-phonon interactions in biased graphitic bilayers' 😅
I feel like everyone gets Moor's law backwards. It was less a theoretical prediction of how fast miniturization would go but was a benchmark to aim for. A self-fulfilling prophecy, we could say.
The whole purpose of Moor's law was to give the varied software/hardware producers a benchmark to aim for over your project's lifespan.
E: In the short term, we're going 3d, just stacking chips and calling it good enough.
Of course NVidia would say the Moores law is dead.
It's because NVidia still uses a single silicon chip to do everything in one place.
Other manufacturers can still keep increasing their computing power because AMD for example has already been splitting their CPUs into smaller chiplets for quite a while and now they're doing the same with their GPUs.
Uhh... is increasing the power and heat while breaking CPUs into more parts really something to brag about? You do realize that what everyone was excited about not that long ago was being able to integrate more and more functionality into a smaller space while using less power. So of course some performance was being left on the table that we can "get back" if we are willing to push more power through and ditch some of the integration for separate parts. At the rate we're going, CPU manufacturers will have to stop integrating GPUs to get more die space, then they'll have to push more and more tasks off to the motherboard chipset, until finally maybe we wind up with separate physical dies for floating point and integer tasks...
@@jeremyandrews3292 Interesting thing to say because AMD is makes BY FAR the most power efficient chips at the moment, and they are using the chiplet design. (Though I have a Steam Deck which I think is a single chip and at the same time is the king in power efficiency so maybe the gains are elsewhere)
So I don't think it's the case of pushing more power and more heat it's just having more transistors overall without huge risk of the whole unit having to be thrown away because of a single defect. Maybe there's some more compartmentalization on the chip into the chiplets and maybe even specialization but not to my limited knowledge, not yet anyway.
Basically instead of doubling the ammount of transistors on a 4 squared cm chip they just double the ammount of transistors by making the chip bigger and they don't suffer from the single point faliure during manufacturing. That's what I'm thinking anyway
Splitting CPUs into chipsets. only helps with manufacturing costs. Although AMD has been desperately working to bring their experience in chiplet design to the GPU manofacturing, they have not been able to do so expect for the I/0 controllers part of the silicon, which does saves a lot of money, not having to spend engineering time shrinking the design of something that could be handled on a different chip.
For now chiplet design has only been possible on the CPU manufacturing because it is easier for CPUs to run in parallel.
The reason chiplet design reduces costs is that during lithorgraphy, printing errors are possible, having a monolithic die, means if you have a printing error, you wasted a lot of silicon. Also chiplet design allows modularity, so you can decide what to do with those chiplets better, do sell a lot of cheap 8 cores 8 threads CPUs or combine them to sell a 32 core 32 thread monster.
Chiplets design does not help in keeping the Moor's law alive, problem here is, like. mentioned in the video, we are getting too close to the side of an atom, at those sizes we start to see quantum tunneling, so an electron can jump to the other side of a gate( the door turning off a transistor) there for preventing any computational calculation.
Moor's law also stated that the number of transistors doubled every two years, yet the production costs were cut in half. And if we consider that second part, that's were Moor's law has been dead for years and years, the production costs have been getting more expensive each generation because is becoming harder and harder to print wafers, chiplet design helps a little bit to mitigate this but in no way allows manufacturing costs to be halved, same for 3D stacking, AMD has been experimenting and researching 3D stacking for years, but manufacturing is just too expensive.
Single silicon atom, is 0.2 nanometers, hour chips are capable of 2, which is means have probably 3 generation before things fall apart unless 3D stacking becomes a thing.
A different semiconductores is great, but to create the labs and machines to mass production and for engineers to figure out their designs that not only are equivalent in performance but are actually better at it, can take decades.
So yeah Moor's Law is dead, and it has been for a while.
As 2 others essentially mentioned already, Moore's Law never addressed the doubling of speed by adding chips... it was solely speaking to the fact that things were getting smaller and faster.
Of course, there is obviously a finite limit to this process so Moore must have known it wasn't really ever capable of being a "LAW"... only a temporary observation until we get to the point we've been at for the last 5-10 years now... SPEED INCREASES are almost non-existent and certainly nowhere near doubling every 2 years... Here's an article from 24 years ago (almost) that says Intel is claiming they are going to be the first to release a 1Ghz processor that year. (Obviously we don't have TERAHERTZ chips today... hell, we don't even have any 6Ghz chips available for retail purchase at anything close to affordable, barring insane overclocking and water cooling tricks): www.zdnet.com/article/1-ghz-intel-claims-it-was-first/
@@oohwha My point was that real performance still does follow roughly Moore's law despite the chips keeping about the same Ghz speed for many years now.
Paralelizing is being properly used by virtually all applications now and the number of cores can be increased arbitrarily if you have chiplets (on a single chip you hit a wall where a single fault is so likely to occur on the chip to make it non-viable as a design)
Stacking of cashe on top of the logic in the "3D" is apparently the thing to do now INSTEAD of just adding more cores because that's where the bottleneck is.
If you had a 100Ghz single core CPU (which should be possible on Graphene actually) it would be nice for a VERY narrow usecase but for gaming (which is the driver for performance let's face it) the 3D chips with large and fast cashe would outperform it simply because of the memory read and write bottleneck.
So you need to carry over the "anciliary" advancements which still follow the Moores Law in order to introduce the new technology anyway. That's why I would personally count these improvements in Moores Law especially because they roughly follow that rate of improvement of overall performance.
Very promising. Of course we will have to wait to see if it can be scaled up to industrial production but I can only imagine that it will, after all modern semiconductor industry is also extremely delicate and sophisticated.
Luis Aldamiz
Why you need to wait, how old are you ?
Don't wait, that is never any productive, why you need quantum computing in the near future, are you able to emulate it now, run it on today hardware ?
@@lucasrem - WTF?!
I don't think quantum computing will ever live to its promise, I'm very skeptical of all these pseudo-science passing as "science", "The Emperor's New Clothes" of our day, not just quantum computing but even more so like nuclear fusion. Now, said that, the challenge to make computers more efficient, either by extending Moore's Law with new materials like this one of by developing new approaches to computing, is probably important and I respect that.
I'm 55, but when I say "we will have to wait" I mean "we" as Humanity, not you and me.
Your my go to science teacher... you explain things like a favorite professor... the one the kids all love
Thanks. Takes me back a few years where we waited for the next device to catch up to what we wanted yesterday. The ol' hurry up and wait routine.😁
Thank you for this video. I learn so much from your channel
It's a hopeful material! Thank you Ms Sabine ❤ Danke!
Nice overview of a very recent result, but imo it has to be noted that majority of the authors are from Tianjin University, China, so GeorgiaTech is not the main contributor here
Thumbs up, Sabine, and I'm glad that I'm already subscribed.
semiconductors - it's not so much the current that gets them to/from their relatively (non)conductive states, but more so voltage, e.g. bias voltage. And this applies even more so for Field Effect Transistors (FETs) and Insulated Gate Field Effect Transistors (IGFETs). Though regular transistors work more as current multipliers, it's still that bias voltage that gets them to conductive state (likewise for diodes). FETs and ITFETs, however, work even more so like voltage controlled devices - like ye olde vacuum tube grid elements - than conventional transistors. So, even though its the current that transistors multiply, it's still that initial bias voltage that gets base-emitter junction up to it's (relatively) conductive state.
It's now written in the errata. Thanks for your explanation!
Thank you very much Dr Hossenfelder. Good video about SEMI. I enjoy listening to your videos and learn. You make it interesting! 🤓
i loved the way you presented that. I am a subscriber.
Why did you leave out the other half of the research team that developed this? It was Georgia Tech along with Tianjin University in China.
It was really Chinese researchers at Tianjin University and Chinese researchers at Georgia Tech.
Doping with boron and nitrogen directly could work too. The junctions between boron carbide and carbon nitride with a graphene conductor already show similar properties when printed using a standard inkjet printer (nowhere near 1nm resolution of course) but sufficient to start a new industry of cheap, rapid circuit development that could eventually get there.
Apply gold so it has an atom supported by 2 nitrogen atoms on the hexagonal ring and 3 boron atoms on the ring above and you have a quantum transistor that can shift states by absorbing light using a phenomenon called tunneling. 😮
There is just so much momentum behind Silicon. The promise would have to be overwhelmingly amazing to blossom beyond niche applications into something a TSMC or Intel (or NewCo) would value enough to implement at scale. Moore’s Law is a technical law…that is really a law of economics.
The very first blue LEDs that I know of used SiC - first approach to obtaining the necessary bandgaps. I have 10 of them in some Christmas tree light strings I made several decades ago. BTW, I would use the word "voltage" rather than current in your description of how semiconductors work. Voltage/potential difference is what gets electrons (current) flowing.
Thank you for the video.
The USA is making it as difficult as possible for China to buy machines from the Netherlands for microchips with 2-nanometer transistor architecture... soon they will have their own machinery and there will be a lot of new developments in this field.
Fascinating! I hope we can see it in the market soon! 😃
Thanks, Sabine!!!
Stay safe there with your family! 🖖😊
The issue is that legacy chips are still the backbone of so much. When this gets matured they will be present in newer stuff sure but older platforms, in a military sense especially where old is writing something, do not rely on such items. There needs to be a way to assure Taiwan and yet dominate newer chips too. Science it is!
Graphene actually *is* a semiconductor. The problem is that there's no energy gap without a PN junction. If you can't dope the crystal (that is, introduce impurities that carry extra electrons or holes into the crystal lattice) then you can't create a PN junction.
Very good. What kinds of dopants would you put into the graphene? What percentage of impurities would you put in to form the PN junction?
@@JoshtMoody That's the problem. Graphene's unique physical properties arise from its unique crystalline structure, not the carbon atoms themselves. You can't dope it without disrupting the structure.
@@GoSlash27all dopants are imperfections that effect the structure, see boron-graphene and graphene oxide
Its not a semiconductor, it's a semi-metal
@@eoingriffin3517half empty = half full?
I believe Sabine is mistaken about transistor size at around 0:40. The process size isn't the size of the transistor but rather a loose approximation the the smallest feature size that can be made using the technology they have (a pixel if you will). Different manufacturers use different standards so one company saying they can do 4nm may not be doing better than another company that says they can do 5nm. I couldn't find the actual size of whole transistors, but looking at transistor densities on Wikipedia, I saw that the latest and greatest was about 125M transistors per mm^2 e.g. Nvidia AD102 GPU ; this works out as transistors having a separation of about 90nm; we might guestimate that the transistors themselves would be perhaps half of that.
Astounding breakdown of graphene's potential in transcending current semiconductor limits! 🌟
I might do research about this in my PhD if I succeed on my doctorate exam (microelectronics) Graphene and finfet transistors have been an interesting research subject for years
Go for it! If it excites you, then dump your effort into your passion! I have a PhD in accelerator physics from UCLA in 2014 and I love my job!
@@JoshtMoody thanks for the encouragement I appreciate it
Biology >> Physics
As someone who studied graphene in their PhD I would push you more toward gallium nitride and silicon carbide. Graphene is great in theory but even now there are only two industrial scale applications (these semiconductors and batteries). But if you find graphene interesting, go for it.
To pass my exam, I was in several groups with one person for a specific subject. We worked on a few problems in every section of upper level UG and beginning level grad books.
We worked for a year.
I worked separately to get my lower level class books and solved every homework problem in them. I still remember a couple where I never did solve them…
A very interesting Episode. Turning novel problem solving into a much better 'second generation' of standard computer chips. You shows are very well produced and informative, and Sabine is a terrific presenter. I have been listening since she wore glasses LOL Subscribed for a LONG time. Cheers Los Angeles
a very good update!
I did come up with a design for architecture that had small, relatively slow modules that were connected in 3D. I was on acid at the time but, given a material that could do this in a 3 dimensional lattice, it would work. The secret is synchronisation so that everything doesn't try to happen at once.
Yeah, what about a switch that can handle multiple signals at the same time wired into a 3D matrix? Oh, we already have human brains…
Well if brains can do it, we can also do it!
Pardon the confession. That’s something like how I manage my telepathic network. But it’s 5D, at least.
@@andoreanesnomeo1706 5D gives much more bandwidth than 3D. 😊
Have you heard of the Anton supercomputer? It's quite a typical approach, to use 3D positioning to reduce routing.
@@vitalyl1327 I haven’t heard of the Anton super computer. I tried to get into super computers in grad school back in the 80s, but went into more theoretical approaches to physics. I think my brain was more like Dr Hossenfelders.
Chip cooling could be done with bimetallic junctions, these would cool the units as current flows between them.
Okay, THAT finally made me Subscribe.. Great CZcams Work Madame H. Thumbs up for this Video.
That research is pretty SiC. Now if only they made it potassium-doped. I hope it ends up being viable to scale though. Can't help but wonder how it behaves with additional magic angle layers of graphene.
Whoa! I just noticed. Sabine, congratulations on serving over a million subscribers!
She needed to have a million members, then she could finance her experiments on superdeterminism.
The cooling issue is important for consumers because a lot of computers generate too much heat. A laptop burns your lap and during the summer a hot computer adds to the heat inside of your home. The heat also makes a potential fire issue as well.
I can’t wait to buy a new laptop with a processor made of this stuff and a solid state battery
I like the new short daily videos Sabina
Bet it still doesn’t cure the blue screen of death.
Top tip - swearing at the computer makes it run faster
I believe something like that will become possible, because it fits with Ray Kurzweil's long-term prediction of how when one technology paradigm runs out of steam another paradigm replaces it and continues the exponential trend. Such as relays, vacuum tubes, transistors, integrated silicon circuits and then in the future maybe things like graphene and/or photonic microchips.
Diamond is excellent at thermal conduction as well. It can be applied with vapor deposition.
Reminds me of Michael Crichton's book Congo (1980), wherein a tech company tries to scoop up an African mine that has type IIB diamonds, which are doped with boron, making them semiconductors. Diamonds also have a thermal conductivity >5x higher than copper and close to 20x the thermal conductivity of silicon, meaning they can disperse extreme amounts of heat.
I go to GA Tech and am a EE that does SoC design. This is so cool
You can use fibre optics for a very fast computer. Add to coherent light sources to produce logic 1. In crease the path length of one light source to produce destructive interference resulting in logic 0.
Sehr schön und gleich zum Punkt!
Or you can use superconductors to reduce heating to zero. I think, at some point of coefficient "performance/energy", cooling chip to superconductors' temperature will be more effective than trying throw temperature away. Of course if chip made from some of high-temperature superC.
great breakdown
Nice ❤ Thank you :)
The company I work for, Wholesale Graphene, manufactures graphene for structural purposes. It is great stuff, paradigm changing. We have also grown graphene on silicon carbide. I have some in my lab.
Unfortunately whenever I hear some amazing improvement that is obtained by graphene I just remember what other amazing improvements that did not translate to actual usable products. Let's hope this can be usable in the next decade.
Weirdly even the cheap ones you can do at home (like Technology Connections' flash graphene-reinforced polymers) don't seem to be common on shelves yet. Even in premium segments. Not sure what's up tbh
I used to work in an Intel fab in photolithography and I approve this message.
Maybe a conical / open screw layout for transistors could work
I just had a discussion about this a few days ago.
I like the conclusion. ;)
I love your videos! Thanks for explaining to mortals so well
AFAIK, the current flow on one plane direction of Graphene can be controlled with an electric field in the orthogonal direction. (or something like that ...)
Anything graphene sounds like a massive game changer, but rarely practical when it comes to production. :( I truly hope the technique in this paper can translate over into current semiconductor fab processes… even just for older ones at first.
Maybe high power applications?
We'll probably get a hell of a lot better if we can prove that graphene based computing is that much better. The problem with graphene isn't that we can't make it, it's that it's expensive to make a lot of it and so funding for research isn't super easy.
If this is another gallium transistor video....right, time to watch.
Sometimes ( many times) I feel a kind of ignorant regarding that subject… but this lady make me understand! She is great!
Two things,
1) Technically Moore's law would still be dead with 3d stacking. Moore's law was addressing the number of transistors in a given area of a silicon wafer. If you're stacking wafers, you're not increasing the density, you're increasing the area. If you move to a material other than silicon, then Moore's law doesn't apply anymore either.
2) Something I didn't see mentioned is that the distance between nuclei in a silicon crystal is roughly 0.2 nm. That's the fundamental limit to the lower threshold of transistor size, except that no it's not, because transistors aren't pure silicon, they have to be doped with, traditionally, either boron or phosphorous to make either P type or N type transistors. If you imagine a transistor that is 1 nm across, that would be a square area "crystal" of 25 silicon atoms, with some replaced by the doping material. 25, atoms. Even if we could wave a magic wand and ignore the quantum tunneling problems, there just aren't that many atoms remaining per transistor to remove and shrink the size. And it's not like those are going to get easier, we're at a point where economics will come into play where it's just too expensive to shrink them any further.
Since quantum computers are 1) not a replacement for traditional transistor computers, and 2) mostly hype without substance, we're rapidly approaching the fundamental limits of computing power getting cheaper over time. There will still be some gains as we can improve the yield and efficiency of existing tech, but we're right next to the end of exponential gains.
Thanks!
Thank you from the entire team!
Spintronics is really interesting!
didn't you mean voltage, not current? if a DUT isn't conducting then no current is flowing.
Ack, you're right of course! Sorry about that.
transistors are great but interconnects are just as important. Even the "size" that you're told that transistors are isn't really accurate; they still have to shrink and find compatible interconnects between the elements
Amazing when it comes to grey friction present in the dynamics of the light visible wavelengths bound to the surface, I would consider spin concentration dynamics of atmospheric sized atoms and molecules as a baselined issue in computing for the future 🙏❤️
I love the "honey/kəʊm/" subtitle at 2:26. Not to be confused with a honey/koʊm/.
I want to warn everyone that most technological breakthroughs are just hype. For example, a lot of companies are reporting the outcome of using AI, and they achieve like 5% productivity improvement. That is still an improvement, but it is not the 100% promised.
100% has not been promised, nor has any specific number. And no promises have been made for current AI systems. For example OpenAI made no claims about how chatGPT-3.5 could be used in businesses. Businesses realised it might be useful themselves and started testing it. OpenAI then responded by trying to make it easier for businesses to do this, but again made no claims about how good it would be (beyond "better" and stating that they hoped it could be used in the ways businesses wanted).
The "promise" (which I mean in the sense of "shows promise" rather than the industry making such a promise) of AI is that it will continue to develop and improve, potentially very rapidly, and that these improvements will lead, in maybe 10-20 years, to massive reworks of many industries' operations.
No one in the industry is under any delusions that today's AI should be able to revolutionise anything (though arguably it is already revolutionising protein research and drug discovery), so testing showing that it doesn't isn't surprising. The point is that AI seems to be progressing rapidly, so we need to prepare for the effects it may have. Not that those effects are already here. You're disproving a claim that the industry has not made. And as Sabine said, from prototype to full deployment is a very long path. It's been less than 18 months since AI fully entered the public conversation with the advent of chatGPT. And since then research and discoveries have exploded, not slowed as one would expect from something that is just hype.
Depends on the type of companies reporting. Recommendation systems are based on machine learning and falls under the AI category, and it has been significant for sites like Amazon and CZcams.
Any single thing doing 5% is a lot IMO.
I was always on team SiC (and diamond) when it comes to next-gen semiconductor. I don't think graphene is competitive as people haven't figured out the basic band gap problem. This video surprised me quite a bit at first. Then it became really satisfying when it is revealed that SiC is also involved. I think the problem is that SiC is really good for high power applications, and there has not been a real candidate for computing.
This is pretty exciting news
Graphene and silicon carbide. No chance of running out of the raw materials!
Gallium di-phosphate carbide whe applied to a diamond Substrate acts like a super charged uni polar conductor. Everything Sabine mentioned is possible with gallium di-phosphate
The transistors are not actually ~3nm or so. That number loosely relates to the 'half pitch' but frankly that's breaking down too. Usually the actual pitch is considerably larger in the several tens of nm and the transistors are made from multiple fins of a finFET. They are much much larger than the 3nm or whatever marketing claim.
This is a fairly common error when people talk about modern mosfets.
Peltier cooling can be used in stacking semiconductors.
Is this another star trek technology becoming reality? Everyone thought it was absurd when those starfleet engineers were using crystals in computers.
This is great news!
Georgia Tech in the news! My buddy went to grad school there
Go Georgia Institute of Technology! This is a great testament to the reach of good science! You don’t need to be at an MIT or Oxford to do important research.
Georgia Tech has been an academic research powerhouse for almost as long as MIT has. It's not at all surprising if a technological breakthrough were to happen there.
@@johncate9541 Absolutely!
It was not just Georgia tech researchers. They were collaborating with researchers in Tianjin University in China. Funny how that got left out of the story
@@hdhdhshscbxhdh4195 And the fact that only the last three authors were from Georgia Tech, the first like 5 authors were all from Tianjin University.
GT has a 20% admit rate for undergrad, the most non-medical research expenditure of all US universities, and consistently is in the top 100 for world rankings. Not quite the same as if this came out of a state flagship.
I remember simulating devices with ferroelectric materials like barium titanate and potassium sodium titante materials as spacers and emitters and collector for transistors, I did observe better transfer characteristics but ultimately existing production industry defines how any new technology can be feasible, pretty sad if you think about it since many research papers I've read showed far more remarkable results!
Arrays of those (barium strontium tantalate) were used in earlier thermal image devices. The sensor chips were incredibly well made. One was soaking in water inside the imager for over s decade and though the read out electronics were total scrap, the imaging chip itself still worked.😮 These were made by Raytheon and used in firefighter cameras, but the module itself was a surplus heat seeker camera module from the Gulf War. Go figure.
And when you stack them, how are you going to get the heat out so it does not melt?
A) Carbon nanotubes can be consucting or semiconducting. B) in transistors you can make them conducting by applying a voltage ...
Surface graphitization of Silicon carbide has been known for decades. I read papers during PhD 15 years ago.
Producers of thermionic valves: "Transistors are rather unappealing because it will require many changes - what will we do with all the glass blowers?"
I suppose that to get an even tinier component, we need a
Semi-hemi-demi-conductor and a small orchestra.
Yeah Graphene the supermaterial everyone was talking about and noone uses because its basically carbon as a metal which only had the advantage of being lighter while thousands of times more expensive than compareable conductivity metals. It would not surprise me if someone else now tries to simply steam or anodize a layer of metal onto silicon to achieve the same thing without the need of producing graphene.
Very cool
Nvidia: "Moore's law is dead"
AMD: "Hold my beer"
Thankfully AMD took a chiplet approach to CPU design in 2017 or we'd still be ground to a halt with performance gains as we were from 2012-2017. Technically that allowed for increased transistor count, just not in the same surface area. More design breakthroughs like that are the future, and i suspect we'll have cuboid CPU's before my time's up.
Hexagons really are the Bestagons
Could it be that Moore's law on transistors isn't about transistors but the rate of our markets and tech to evolve due to demand. Another interesting issue is that when there closer gaps between transistors will we hit an issue where they can arc between each other, I also often thought that at the size of atoms would a new emergent ability occur with transistors where we could manipulate those atoms and/ or fields
They're already having these problems, that's why we need a new material.
No. Moore's Law originally was the doubling of transistors in 1 year not two years. In 1975 after 11 years Moore adjusted the doubling to 18 months and then over the next decade it was changed again to every two years.
Shrinking transistors is the most obvious way to reduce chip size, but I think there should also be more research in trying to redesign the logic circuits so they need less transistors. There is probably a way to redesign the full adder to use less transistors. Implementing floating point math in software might be easier, to pick a few examples.
Radix-4 floating-point arithmetic can be done quicker than radix-2. Radix-60 arithmetic is more accurate than radix-2 in long computations.
Dr. John Gustafson, of Gustafson's Law, has published a book targeted at exactly this problem. Entitled The End of Error, it describes a new and better way to do floating point...or not do most of it, and far more efficiently.
Circuit optimization has been happening for 20+ years, as the fab cost have been increasing rapidly.
Interesting, happy to hear really thats pure carbon based materials are getting spotlight especially graphene and CNTs. I am particularly fascinated about their structural marvelousness. Did my MEng thesis on armchair CNTs and its hust incredible. Can't wait to see better and more affordable ways of manufacturing them.