Graphene Interconnects Aim to Give Moore's Law New Life
spectrum.ieee.org
spectrum.ieee.org
Secondly, process fab is VERY conservative. There’s numerous amazing ferroelectrics that you can grow tons of that would absolutely spank NAND flash. However, they’re not silicon fab, so nobody makes them.
So why doesn't somebody new start making them and put all the current flash producers out of business?
Is there actually a special property of scotch tape that makes it the ideal candidate over some more specialized industrial adhesive? Or are these references to scotch tape generally just references to the fact that you _can_ use scotch tape like the original graphene experiment?
Graphene has incredible properties as a structural material too but so far producing it at that scale and making it behave properly in things like composites has been very hard. But the physics says once we get it to work we have composites many times stronger than steel or materials like Kevlar.
And we have been able to produce graphene around 2004 I believe, so we are going soon to cross that threshold.
I'll give you fifteen minutes to call me back.
/Jerry Maquire out
Incredible
But I’m already thinking about light CPUs that use light instead of electricity for computation. Of course I don’t fully know how it works but it seems to be lower power and the next iteration of computation I guess before we get to room temp quantum computers.
If you get enough other benefits from going up from 2nm features to e.g. 200nm UV-C photons, then you may still choose to do so.
If you have a system clock running at 3 GHz, the speed of light limits your causal distance to just under 10cm per clock cycle. CPUs are already close to that for size and frequency, but let's say you're taking a 1 cm by 1cm silicon chip on a 2nm feature size process and replacing it like-for-like with a photonic chip with light that limits it to 200nm features — now it's 1m by 1m and can't go faster than 300 MHz, likely a lot less.
This doesn't mean it is useless — for example, there's a hope that it will reduce energy use, which is directly useful all by itself, but also means it may be sensible to move to a fully 3D structure which silicon can't really manage because of the thermal issues. Going from 2D to 3D helps a lot, might allow that 1m by 1m by 200nm (*2 thickness for insulation) sheet to be compacted to a 7.4mm cube, which then doesn't need to be slowed all the way down to 300 MHz due to causality.
All of these are much lower density than magnetic (hard drives) or electrical (RAM or flash).
I've not heard of RAM or flash having thermal issues (though I'm sure it would happen eventually), but that suggests 3D stacking is easier for storage, which would remove that potential advantage of optical.
One thing I've seen suggested for optical computing is to create optical elements (e.g. lenses, holograms) that represent the same transformations as a layer of a neural net; they're not at all space-efficient and any changes to the network requires basically replacing the whole thing, and you can do a similar thing with a suitable network of hard-wired resistors and transistors, but they're an interesting idea that I see come up very occasionally.
* Some people have weird ideas.
Among other reasons, you can create an electronic transistor in silicon by using an electrical signal to open and close a gate.
You can't really do this with light, light beams just pass through each other. And the kind of light-carrying media that can be affected by the presence of a control beam respond much slower and less effectively than doped silicon responds to voltage.
Nothing beats CMOS transistors in density.
You're clearly not using enough power for the Schwinger effect. (More seriously, there are other non-linear effects in non-vacuum optical media.)
https://en.wikipedia.org/wiki/Moore%27s_second_law
It seems likely that we're relatively close to the point where it will no longer be economical to push the limits here. It's unlikely that even the entire world working together would want to spend more than $1T for a single fab, which Rock's law suggests is less than 20 years away.
It seems likely that we're relatively close to the point where it will no longer be economical to push the limits here. It's unlikely that even the entire world working together would want to spend more than $1T for a single fab, which Rock's law suggests is less than 20 years away.
Twenty years is a long time. I don't think people in 2025 could have predicted out needs for chips back in 2005.
If each company is worth $20 trillion in 20 years, they can easily raise $1 trillion together by selling some shares or using their shares as leverage or just straight up using their cash flow. I'm simplifying things by ignoring inflation, but you get the point.
The bottom line is, if capitalism thinks a $1 trillion fab will produce more than $1 trillion in value, it will happen.
Which is to say, if Rock's law continues to hold, it doesn't matter if some global consortium can pull together $1T for a new fab; can they pull together $2T four years later? And $4T four years after that? And $8T? $16T? To say that a doubling at this rate is sustainable is to suggest that you more than double the value at each step. At some point this can clearly not be the case, unless you want to posit a world where going from one process node to the next literally doubles the entire productive output of the human race.
Absent some unforeseeable technological breakthrough, at some point it has to slow down, either slowly, or drastically, or otherwise stop altogether. And for anyone who's currently middle age or younger, it's currently projected to happen in your lifetime.
TSMC has always had N2 on its roadmap long before the AI boom.
Also, back side power delivery + new materials like Ru will keep interconnect roadmap going for a while.
Articles like this read nothing more than fluff pieces.
No doubt, massive technological advancement if it can me massively produced.
But we’ve been waiting 20-years and still haven’t realized the benefits.
You can even make graphene at home with a blender.