First 3D Nanotube and RRAM ICs Come Out of Foundry(spectrum.ieee.org)
spectrum.ieee.org
First 3D Nanotube and RRAM ICs Come Out of Foundry
https://spectrum.ieee.org/nanoclast/semiconductors/devices/first-3d-nanotube-and-rram-ics-come-out-of-foundry
2 comments
3D circuits make a lot of sense. I'd be interested to know how they plan to cool this though. My understand was that the in current chips, the 3rd dimension is largely used for heat dissipation, and that heat is the primary constraint on faster processors.
This is great, yes. Unfortunately, they don't address the economics of foundry latency though. Monolithic 3D means way more layers which means way more production steps, which means it takes longer for the wafer to make it through the factory. We're already talking about production latencies on the order of months as it is today... so this is clearly cool tech, but scaling further may be difficult.
Does this mean that the news of Moore's law demise are greatly exaggerated?
It seems likely that Moore’s Law will someday be resurrected, perhaps in a weaker form, but I wouldn't bet heavily on any particular invention being the one that resurrects it.
I say it seems likely because Landauer's limit works out to about 0.003 attojoules per bit erasure at room temperature, while current top-efficiency processors use on the order of 100–200 picojoules per instruction — say, about 1 picojoule per bit erasure. So we're about 10 orders of magnitude from Landauer's limit, even before we switch to reversible computation, or drop Landauer's limit by a factor of 75 by operating our computers at the temperature of the cosmic background radiation. Other fundamental limits on computation (Lloyd's "ultimate laptop": https://arxiv.org/abs/quant-ph/9908043) are even further off.
In 2D, though, we're actually sort of close, in the sense that the transistors TSMC and Samsung are mass-producing are only about five orders of magnitude larger than the somewhat fundamental limit of single-atom transistors; https://news.ycombinator.com/item?id=20273007 says silicon atoms in an unstrained lattice are 0.235 nm apart, so a 10-nm-wide gate is 40 atoms across. Single-atom transistors have been working for decades in the lab (IBM Almaden, maybe?) so we know there aren't any fundamental limits to computation in between here and there.
I say it seems likely because Landauer's limit works out to about 0.003 attojoules per bit erasure at room temperature, while current top-efficiency processors use on the order of 100–200 picojoules per instruction — say, about 1 picojoule per bit erasure. So we're about 10 orders of magnitude from Landauer's limit, even before we switch to reversible computation, or drop Landauer's limit by a factor of 75 by operating our computers at the temperature of the cosmic background radiation. Other fundamental limits on computation (Lloyd's "ultimate laptop": https://arxiv.org/abs/quant-ph/9908043) are even further off.
In 2D, though, we're actually sort of close, in the sense that the transistors TSMC and Samsung are mass-producing are only about five orders of magnitude larger than the somewhat fundamental limit of single-atom transistors; https://news.ycombinator.com/item?id=20273007 says silicon atoms in an unstrained lattice are 0.235 nm apart, so a 10-nm-wide gate is 40 atoms across. Single-atom transistors have been working for decades in the lab (IBM Almaden, maybe?) so we know there aren't any fundamental limits to computation in between here and there.
Will just go in a different direction. Gordon Moore himself alludes to advanced packaging / 3D configurations in his original 1965 paper:
"It may prove economical to build large systems out of smaller functions, which are separately packaged and interconnected."
"It may prove economical to build large systems out of smaller functions, which are separately packaged and interconnected."
Moore's law is dead right now. Maybe there will be a successor, maybe not, but at the moment one does well to not assume that we will get another free lunch of processing power soon.
It just slowed down a bit.
transistor density is still improving logarithmic.
Moore's law is about the number of transistors on a chip, not about performance.
transistor density is still improving logarithmic.
Moore's law is about the number of transistors on a chip, not about performance.
> Moore's law is about the number of transistors on a chip, not about performance.
People talking about Moore's law talk about the "more transistors means more performance" aspect. That part hasn't held for a while now. No one cares if you have ten times as many transistors if it doesn't improve performance.
People talking about Moore's law talk about the "more transistors means more performance" aspect. That part hasn't held for a while now. No one cares if you have ten times as many transistors if it doesn't improve performance.
Performance of GPUs has increased quite well over the last 10 years. Number of Cores on CPUs have also increased.
Even if performance didn't improve, that wouldn't make moore's law dead.
You might say that Dennard scaling is dead. But Dennard scaling isn't Moore's law even thought is gets conflated a lot.
Edit:
Compare a GTX 285 to a RTX 2080 ti, the performance has increases more than 9 times over the last decade.
Even if performance didn't improve, that wouldn't make moore's law dead.
You might say that Dennard scaling is dead. But Dennard scaling isn't Moore's law even thought is gets conflated a lot.
Edit:
Compare a GTX 285 to a RTX 2080 ti, the performance has increases more than 9 times over the last decade.
> Number of Cores on CPUs have also increased.
Correct, though number of cores is far harder to use in software than the higher Mhz/Ghz we enjoyed before. Same for GPUs. If you can use them you still enjoy a noticable increase in performance, but not all programs (or rather algorithms) can be changed to benefit from more cores and even less commercial ones.
> You might say that Dennard scaling is dead. But Dennard scaling isn't Moore's law even thought is gets conflated a lot.
Fair enough. I looked this up and the combination of Dennard scaling and Moore's law seems to be called Koomey's law - I've never heard that term before, but it fits the definition of what people usually attribute to Moore's law:
> Jonathan Koomey articulated the trend as follows: "at a fixed computing load, the amount of battery you need will fall by a factor of two every year and a half."
https://en.wikipedia.org/wiki/Koomey%27s_law
Correct, though number of cores is far harder to use in software than the higher Mhz/Ghz we enjoyed before. Same for GPUs. If you can use them you still enjoy a noticable increase in performance, but not all programs (or rather algorithms) can be changed to benefit from more cores and even less commercial ones.
> You might say that Dennard scaling is dead. But Dennard scaling isn't Moore's law even thought is gets conflated a lot.
Fair enough. I looked this up and the combination of Dennard scaling and Moore's law seems to be called Koomey's law - I've never heard that term before, but it fits the definition of what people usually attribute to Moore's law:
> Jonathan Koomey articulated the trend as follows: "at a fixed computing load, the amount of battery you need will fall by a factor of two every year and a half."
https://en.wikipedia.org/wiki/Koomey%27s_law