So you get a ~1300 reduction in clock speed but reduces power consumption by a factor of ~7 million. That still radically improves performance per watt by a still impressive fact of nearly ~5400. The problems shift else where to get 1300 of these chips operating in parallel to get the same equivalent performance (presuming perfect parallel scaling) as well as dealing with higher costs. Granted this is new experimental technology so costs are naturally higher right now but even presuming the same manufacturing costs, there is a need to scale up by a factor 1300 to maintain the equivalent performance due to the clock speed reduction. That is the real barrier for things like HPC and AI to adopt this technology. Scaling has to be at a level far greater than exists today to compensate for the clock speed deficit. However, not to be a total downer those power consumption saving would be still beneficial in the low power embedded space. Not everything needs to be running at multiple Ghz to do its job: the computing space operated between 4 Mhz and 33 Mhz for half a decade on terribly ancient designs four decades ago. These types of transistors would be good for devices that get power off of extremely low power sources. Think hundreds of milliwatts off of solar or motion generators that provide power in bursts (smart watches). Even as things are now with low clocks, there is a niche these could occupy. There are a few other ways to leverage this technology for the power savings. Multichip packaging using the more traditional transistor technologies are used for high speed, shared interconnects between dies and these new transistors used for the compute dies. Imagine a full sized wafer as an interconnect base multiple layers of 2D TFETS dies stacked on top. The base reduction in power is massive so going ultra dense in a fashion like this is feasible but ungodly expensive and challenging to manufacture. 2D TFETS are impressive in the lab and certainly worth watching but also one of those technologies that could always be 5-10 years away as the manufacturing and scaling are resolved. I would imagine that issues like the clock speed reductions are addressed during this phase and costs would certainly have to come down. This just feels like a 'tomorrow' technology as the trade offs to do it today are not feasible.
The equation for transistor power draw is Frequency x Capacitance x Voltage^2, so assuming the new design doesn’t have a higher capacitance or affect clock speeds, the 70% voltage savings would result in a 91% power saving
The reason silicon was selected was as the standard chip material was not just because of its electrical properties, its because it is abundant and therefore cheap. Many chips using other elements have been proposed over the years but they never come to production because of economics. Good luck bringing anything heavier than iron to mass market.
It's the self passivating behavior that made silicon the dominant material. Semiconductor grade silicon, never mind grown and cut wafers, aren't related to the cost of regular sand raw material.
check out intel tera herz. the 70% improvement Ian is talking in this vid is just a joke to keep milking the market with old (and obsolete) silicon tech. to not make actually fast progress so industry can keep milking people's wallets with slow and steady near-0 progress which is good for profit margins and not good for actually making a difference. "new tech" yet basically same results.
The Commodore 64 is a machine featuring a CPU running at roughly 1 MHz, or roughly what those TFETs would deliver at 0.53v Apparently you can do a surprising amount of things with a C64 in realtime regardless of such a restriction. It's a very apropos throwback to "current generation FETs" minus 40 years of steady progress. :- )
@@TheDoomerBloxyeah I recently even read about some generative AI running on C64. Needless to say, it was very slow and only produced something like 16x16 pixel sprites, and was certainly not trained on a C64. But still quite impressive. I'm wondering how quickly this transistor technology could be scaled up in performance. It could perhaps be first utilized in stuff like really low power cell phones, micro controllers and maybe also things like displays (though I don't know how much power their control circuits actually use). Maybe we will eventually need to shift to more and more parallel processing instead of cranking up the clock speed in any case.
@@fintux It is designed roughly in the area you start by describing. You don't need a massive amount of switching speed, if you have a network of decision-trees to traverse - it's usually more handy to not require a lot of power. So really this could already be utilized for voice recognition/modelling "models", implemented as a hardware ASIC of that particular model. And if it was, then it would practically be a robot brain, with ridiculously low energy consumption per neuron traversal, just as in regular fat brains. Conspicuously low clockspeeds on both, too..
@@TheDoomerBlox yeah, I was originally going to mention something about the brain clockspeed, though it's not something you can really define. But individual neurons have maximum firing rates ranging from 1 Hz to 1 kHz depending on the type of the neuron, so I guess that could be considered the maximum clockspeed of a brain. So pretty low, but still manages to process massive amounts of information.
This is awesome!!! I know it is probably a long way away but given what I am understanding it would be very efficient. And it makes total sense: If the source is n-type like and the drain is p-type like, it would make sense that you would need much less energy to connect them.
Hi Ian, Is your news related to the one posted on EurekAlert on 03.07.2024 with the headline -"Scientists discover way to "grow" sub-nanometer-sized transistors Discovering the path to next-generation semiconductors through epitaxial growth of new van der Waals materials" ? Because it's about the same material (molybdenum disulfide). PS. Also published in Nature Nanotechnology Journal...
Great vid👍 Jim Keller is right, however market want’s the compatible legacy design and status quo for as long as possible to avoid adjusting to new foundations due to higher costs and unforseen risks. If this part of the issue would be taken care of better we could see more revolutionary chip designs entering the market instead of good old stuff we r used to. Status quo is indeed comfortable, however it’s an enemy of innovation.
Keep in mind we have companies such as AtomicSemi, and a great awakening in the engineering class of America after COVID showed everyone why global shipping isn't worth relying on. Hopefully, with adequate effort we can revolutionize fab construction, and allow more people to join the innovation circus.
@@Jordan-ru8yf Of course “AMD Ryzen AI 9 HX 370” branding is terrible. Pointing out problems with intel doesn’t negate problems with AMD. It’s not a zero sum game. Also AMD isn’t a foundry, hasn’t been for a while, so not really relevant to the video. Not everything is Intel vs AMD man.
I think switching speed has become less and less of a bottleneck, see eg. domino/dynamic logic. The delays in the wiring have become an increasingly more important part of the effective speed of the circuit.
Slower per node is not bad. 700Mhz is way, way faster than our brains. Though the brain only uses about twenty (20) watts of energy. On the other hand, may intel continue with higher scale general purpose CPUs used in today's laptops.
I wish they would also try to find ways to extract more heat out of traditional FinFet. 8Ghz single-core performance would be amazing for compiling stupid software made in C++, for example.
Ah. I was going to ask what the speed of switching is, but there we go, you answered. Yeah this is not going to be ready for mainstream computation anytime soon. Very interesting though. I imagine this will be snapped up for mobile compute (tiny wearables etc) as soon as they can make it feasible.
aren’t most of the fastest modern transistors 3D though now? it seems like the tradeoff here would be speed vs power consumption unless there’s some way to have the gate surrounding the source & drain like finfets edit: 8:06 yeah that sounds about right but idk if that’s really just about development. seems like a more fundamental limitation with this method.
The zx spectrum was the competitor to the commodore 64, i had an amstrad 486 The problem is the 1nm barrier will we go photonic or carbon nano tubes I know there is 3d transistors and i have seen some other methods There is no chance that quantum computer will go mainstream
Please techpotato, make a video about this: Progress of Tsinghua SSMB EUV Light Source Development Xiujie Deng, Tsinghua University, Beijing, China On behalf of the SSMB Collaboration 2021 EUVL Workshop, Held Online, June 5-10, 2021.
well according to wikipedia tellurium is as rare as platinum and gold :( en.wikipedia.org/wiki/File:Elemental_abundances.svg seems like molybdenum is much more common so maybe we can see more of it being used
0:23 did you really just censor a Intel stock cooler whilst having a Intel data center GPU pillow and a AMD epyc CPU sitting right there in the corner of your frame? You are a strange one.
Please cut back on the transition effects, they are too slow and VERY distracting. Especially the yellow laser things when the composition of the A-roll doesn't change at all, it just looks weird. But it was a very informative video regardless.
According to physics, only deleting bits has to consume energy. Not setting them or using them for calculations. So if we got _there_ I'd already be very happy. No need to break physics altogether for now :D
transistors work by energy buckets - it takes effort to fill up the gate ... we then dump the gate to ground, while theoretical compute is "free", transistor based architectures are not.
@@andytroo a ballistic deflection transistor is still called a transistor. So I get your point about current technology but it's not the transistor's fault. Not the word anyway.
@@theemperorofmankind3739 Your highness mr. emperor sir, judging by your name it seems YOU you are the one with problems. But since you are a young kid that likes manga I'll let it go. However I have a suggestion, learn your English well then you can post in English.
Blink twice if Serve the home is holding you hostage
Nah, Patrick would not do that!.
So you get a ~1300 reduction in clock speed but reduces power consumption by a factor of ~7 million. That still radically improves performance per watt by a still impressive fact of nearly ~5400. The problems shift else where to get 1300 of these chips operating in parallel to get the same equivalent performance (presuming perfect parallel scaling) as well as dealing with higher costs. Granted this is new experimental technology so costs are naturally higher right now but even presuming the same manufacturing costs, there is a need to scale up by a factor 1300 to maintain the equivalent performance due to the clock speed reduction. That is the real barrier for things like HPC and AI to adopt this technology. Scaling has to be at a level far greater than exists today to compensate for the clock speed deficit.
However, not to be a total downer those power consumption saving would be still beneficial in the low power embedded space. Not everything needs to be running at multiple Ghz to do its job: the computing space operated between 4 Mhz and 33 Mhz for half a decade on terribly ancient designs four decades ago. These types of transistors would be good for devices that get power off of extremely low power sources. Think hundreds of milliwatts off of solar or motion generators that provide power in bursts (smart watches). Even as things are now with low clocks, there is a niche these could occupy.
There are a few other ways to leverage this technology for the power savings. Multichip packaging using the more traditional transistor technologies are used for high speed, shared interconnects between dies and these new transistors used for the compute dies. Imagine a full sized wafer as an interconnect base multiple layers of 2D TFETS dies stacked on top. The base reduction in power is massive so going ultra dense in a fashion like this is feasible but ungodly expensive and challenging to manufacture.
2D TFETS are impressive in the lab and certainly worth watching but also one of those technologies that could always be 5-10 years away as the manufacturing and scaling are resolved. I would imagine that issues like the clock speed reductions are addressed during this phase and costs would certainly have to come down. This just feels like a 'tomorrow' technology as the trade offs to do it today are not feasible.
Don't forget things like computers onboard spacecraft. Reducing the power requirements also reduces cooling requirements and thus the launch costs.
The equation for transistor power draw is Frequency x Capacitance x Voltage^2, so assuming the new design doesn’t have a higher capacitance or affect clock speeds, the 70% voltage savings would result in a 91% power saving
The reason silicon was selected was as the standard chip material was not just because of its electrical properties, its because it is abundant and therefore cheap. Many chips using other elements have been proposed over the years but they never come to production because of economics. Good luck bringing anything heavier than iron to mass market.
It's the self passivating behavior that made silicon the dominant material. Semiconductor grade silicon, never mind grown and cut wafers, aren't related to the cost of regular sand raw material.
There must be a lot of research into 2D materials. Just a few hours ago I watched a video about new capacitors using MoS2.
check out intel tera herz. the 70% improvement Ian is talking in this vid is just a joke to keep milking the market with old (and obsolete) silicon tech. to not make actually fast progress so industry can keep milking people's wallets with slow and steady near-0 progress which is good for profit margins and not good for actually making a difference. "new tech" yet basically same results.
Wearing a shirt with commodore 64 name on it and talking about transistors of the future what a twist :D
Don't you mean 'what a switch'
The Commodore 64 is a machine featuring a CPU running at roughly 1 MHz, or roughly what those TFETs would deliver at 0.53v
Apparently you can do a surprising amount of things with a C64 in realtime regardless of such a restriction.
It's a very apropos throwback to "current generation FETs" minus 40 years of steady progress. :- )
@@TheDoomerBloxyeah I recently even read about some generative AI running on C64. Needless to say, it was very slow and only produced something like 16x16 pixel sprites, and was certainly not trained on a C64. But still quite impressive. I'm wondering how quickly this transistor technology could be scaled up in performance. It could perhaps be first utilized in stuff like really low power cell phones, micro controllers and maybe also things like displays (though I don't know how much power their control circuits actually use). Maybe we will eventually need to shift to more and more parallel processing instead of cranking up the clock speed in any case.
@@fintux It is designed roughly in the area you start by describing.
You don't need a massive amount of switching speed, if you have a network of decision-trees to traverse - it's usually more handy to not require a lot of power.
So really this could already be utilized for voice recognition/modelling "models", implemented as a hardware ASIC of that particular model.
And if it was, then it would practically be a robot brain, with ridiculously low energy consumption per neuron traversal, just as in regular fat brains. Conspicuously low clockspeeds on both, too..
@@TheDoomerBlox yeah, I was originally going to mention something about the brain clockspeed, though it's not something you can really define. But individual neurons have maximum firing rates ranging from 1 Hz to 1 kHz depending on the type of the neuron, so I guess that could be considered the maximum clockspeed of a brain. So pretty low, but still manages to process massive amounts of information.
Great episode! Thanks for taking the chance to do something a bit different!
Great in-depth vlog article as usual Ian.
Loving the new studio setup
Looking great bud!
5:42 when someone told me about quantum computing, this is what I thought they were meaning. Quantum tunneling transistors 🎉
This is awesome!!! I know it is probably a long way away but given what I am understanding it would be very efficient. And it makes total sense: If the source is n-type like and the drain is p-type like, it would make sense that you would need much less energy to connect them.
Hi Ian,
Is your news related to the one posted on EurekAlert on 03.07.2024 with the headline -"Scientists discover way to "grow" sub-nanometer-sized transistors Discovering the path to next-generation semiconductors through epitaxial growth of new van der Waals materials" ?
Because it's about the same material (molybdenum disulfide).
PS. Also published in Nature Nanotechnology Journal...
I only clicked on this video, because of the Commodore 64 T-Shirt. But then stayed for the content.
If you could travel back through time, is this Explaining Computers in his younger days?
This was a really good case of Ian’s chemistry background meeting computers and us prospering.
Great vid👍 Jim Keller is right, however market want’s the compatible legacy design and status quo for as long as possible to avoid adjusting to new foundations due to higher costs and unforseen risks. If this part of the issue would be taken care of better we could see more revolutionary chip designs entering the market instead of good old stuff we r used to. Status quo is indeed comfortable, however it’s an enemy of innovation.
Keep in mind we have companies such as AtomicSemi, and a great awakening in the engineering class of America after COVID showed everyone why global shipping isn't worth relying on. Hopefully, with adequate effort we can revolutionize fab construction, and allow more people to join the innovation circus.
Can we get a video on 4DS Memory PCMO area based, Non-Filament interface switching ReRAM?
This was awesome!
THIS is Ian from TTP and THIS is a BILLION transistors!
I already know intel marketing would go crazy with “ultra quantum-era silicon” or something
Not defending Intel, but finger pointing is bad. AMD just randomly put AI in their CPUS lol. Marketing on both side would do what you propose
@@Jordan-ru8yf Of course “AMD Ryzen AI 9 HX 370” branding is terrible. Pointing out problems with intel doesn’t negate problems with AMD. It’s not a zero sum game.
Also AMD isn’t a foundry, hasn’t been for a while, so not really relevant to the video. Not everything is Intel vs AMD man.
How large or small are TFETs compared to MOSFETs?
Could this make SRAM energy efficient ?
why nobody is talking about intel tera herz? is it forbidden topic for those who works with intel?
great information, thanks.
With lower voltages, this technology has potential for much faster switching in the future. Maybe we'll finally get that 10 GHz P4.
I think switching speed has become less and less of a bottleneck, see eg. domino/dynamic logic. The delays in the wiring have become an increasingly more important part of the effective speed of the circuit.
this comment does not scale
Slower per node is not bad. 700Mhz is way, way faster than our brains. Though the brain only uses about twenty (20) watts of energy.
On the other hand, may intel continue with higher scale general purpose CPUs used in today's laptops.
Yes, but the chart said 0.7 MHz, that is 700 KHz. That’s highly likely to improve, but it’s a very long way from 700MHz
I'd be willing to go with slower/lower clock refresh for 70% less power. Definitely.
I wish they would also try to find ways to extract more heat out of traditional FinFet. 8Ghz single-core performance would be amazing for compiling stupid software made in C++, for example.
@@monad_tcp I agree with you. Compiling or database or running FFT. Companies still do work on desk top computers.
@@monad_tcp Ask for better compilers. Most projects have a lot of code that can be compiled in parallel.
oh noooo i thought this the way they were making transistor 😂😂😂😂
tunneling is as old as dinosaur
Ah. I was going to ask what the speed of switching is, but there we go, you answered. Yeah this is not going to be ready for mainstream computation anytime soon. Very interesting though. I imagine this will be snapped up for mobile compute (tiny wearables etc) as soon as they can make it feasible.
Didn't one of the original Acorn ARM CPU chips already meet this challenge of low power by running when it wasn't even lugged in?
Love videos on semiconductors ❤
good editing good production!
aren’t most of the fastest modern transistors 3D though now? it seems like the tradeoff here would be speed vs power consumption unless there’s some way to have the gate surrounding the source & drain like finfets
edit: 8:06 yeah that sounds about right but idk if that’s really just about development. seems like a more fundamental limitation with this method.
This is still a 3D transistor design, but the transistor materials themselves are 2D. So it's still GAA or beyond :)
The zx spectrum was the competitor to the commodore 64, i had an amstrad 486
The problem is the 1nm barrier will we go photonic or carbon nano tubes
I know there is 3d transistors and i have seen some other methods
There is no chance that quantum computer will go mainstream
2:26 that's related to the gain of amplification of the transistor, are they still 40000 ?
Jim K approves this message.
Please techpotato, make a video about this:
Progress of Tsinghua SSMB EUV
Light Source Development
Xiujie Deng, Tsinghua University, Beijing, China
On behalf of the SSMB Collaboration
2021 EUVL Workshop, Held Online, June 5-10, 2021.
0:06 it's always been price, performance then power for everyone I know.
donno im factoring in power more these days. still price to performance tho
power is heat and i cant have diminishing returns
holidays end. dr. techtechpotato knocks.
Mosfet really seems to be the x86 on that industry.
I've had light switches go into an in-between state. Would not recommend.
I would wager almost anything that for vast majority of people it's performance, price, aesthetics
Gotta get that RGB
STH future collab?
Patrick's always down to collab. Just need the right topic. But he has just started a family a few weeks back
@@TechTechPotatoI can imagine little Partick assembling miniPC lego blocks. One day clustering them....
@@TechTechPotato Wow! That is awesome news happy for him!
well according to wikipedia tellurium is as rare as platinum and gold :(
en.wikipedia.org/wiki/File:Elemental_abundances.svg
seems like molybdenum is much more common so maybe we can see more of it being used
Not sure that matters with the amounts needed.
Molybdenum disulfide is also used as a lubricant.
You forgot the red arrows in the thumbnail. You got the "weird" face, but not the arrows. ;)
0:23 did you really just censor a Intel stock cooler whilst having a Intel data center GPU pillow and a AMD epyc CPU sitting right there in the corner of your frame?
You are a strange one.
My guess is that it is stock footage that was already censored.
@@jeroenvangoch8886 i hope so, it would be so strange if it wasnt
@@deltacx1059Bro of course it's a stock shot, just like the one literally right before it.
Full marks for C64
Please cut back on the transition effects, they are too slow and VERY distracting. Especially the yellow laser things when the composition of the A-roll doesn't change at all, it just looks weird. But it was a very informative video regardless.
At least he doesn thave some loud annoying music playing. I liked it
According to physics, only deleting bits has to consume energy. Not setting them or using them for calculations. So if we got _there_ I'd already be very happy. No need to break physics altogether for now :D
transistors work by energy buckets - it takes effort to fill up the gate ... we then dump the gate to ground, while theoretical compute is "free", transistor based architectures are not.
@@andytroo a ballistic deflection transistor is still called a transistor. So I get your point about current technology but it's not the transistor's fault. Not the word anyway.
So nothing concrete or expected for the next few years. Why even talk about it?
Because research is fun!
@@TechTechPotato But technologies that never gout out of the lab is not fun at all, and I've seen a lot of the in the last 20 years.
@@profounddamas That seems like a you problem.
Covering the predecessors to what will be the next generation of technology is interesting.
@@theemperorofmankind3739 Your highness mr. emperor sir, judging by your name it seems YOU you are the one with problems. But since you are a young kid that likes manga I'll let it go. However I have a suggestion, learn your English well then you can post in English.
Keep the freaking slide up otherwise this is just some marketing stunt
fourth
first!
third
second