The other option is superconducting circuits, which work in labs but also have to many disadvantages to be practical yet. Eventually cold superconducting reversible computers could bypass energy issues, but that tech appears to be far.
What about just replacing the copper wire inside a conventional CMOS chip with a superductor? It took some searching, but I managed to find a paper titled Cryogenically Cooled CMOS which talks about the benefits and feasibility of doing this. Quoting from the relevant section:
If lower interconnect resistance improves performance, the use of ‘zero-resistance’ superconductors should provide the
ultimate in performance improvement. Unfortunately, although performance improvements would be expected, they would not
be as great as the simplistic statement above
suggests. Furthermore, several technical
obstacles remain before high-temperature
superconductors (HTS) can be effectively
integrated with VLSI technology.
Actually, as we will see, the resistance
of superconducting films is not truly zero,
except in the limits of zero frequency or
zero temperature. Nevertheless, at 77 K
and 1 GHz, measurements on patterned
YBa2Cu3O7-x (YBCO) films have already
demonstrated surface resistances one to two
orders of magnitude below those for Cu
under the same conditions. Theoretical
predictions for YBCO suggest four orders
of magnitude would be possible. Unfortunately, good-quality (epitaxial) YBCO films
grow best on perovskite substrates having
high dielectric constants. Lanthanum aluminate (LaAlO3), which is a popular substrate for HTS microwave circuits, has a
relative dielectric constant of 25. Assuming
the same interconnect geometry, this makes
all capacitances more than 6× greater than
would be the case for a SiO2 dielectric.
Thus, some of the low-resistance benefits
of HTS films are cancelled by the high dielectric constants of their associated substrates.
So it looks like there’s no fundamental reason why it couldn’t be done, just a matter of finding the right substrate material and solving other engineering problems.
What about just replacing the copper wire inside a conventional CMOS chip with a superductor?
That is the type of tech I was referring to by superconducting circuits as precursor to full reversible. From what I understand, if you chill everything down then you also change resistance in the semiconductor along with all the other properties, so it probably isn’t as easy as just replacing the copper wires.
A room temperature superconductor circuit breakthrough is one of the main wild cards over the next decade or so. Cryogenic cooling is pretty impractical for mainstream computing.
So it looks like there’s no fundamental reason why it couldn’t be done, just a matter of finding the right substrate material and solving other engineering problems.
Yeah, its just a question of timetables. If it’s decades away, we have a longer period of stalled moore’s law during which AGI will slowly surpass the brain, rather than rapidly.
From what I understand, if you chill everything down then you also change resistance in the semiconductor along with all the other properties, so it probably isn’t as easy as just replacing the copper wires.
From the sources I’ve read, there aren’t any major issues running CMOS at 77 K, you only run into problems at lower temperatures, less than 40 K. I guess people aren’t seriously trying this because it’s probably not much harder to go directly to full superconducting computers (i.e., with logic gates made out of superconductors as well) which offers a lot more benefits. Here is an article about a major IARPA project pursuing that. It doesn’t seem safe to assume that we’ll get AGI before we get superconducting computers. Do you disagree, if so can you explain why?
There was similar interest in superconducting chips about a decade ago which was pretty much the same story—DARPA/IARPA spearheading research, major customer would be US intelligence.
The 500 gigaflops per watt figure is about 100 times more computation/watt than on a current GPU—which is useful because it shows that about 99% of GPU energy cost is interconnect/wiring.
In terms of viability and impact, it is still uncertain how much funding superconducting circuits will require to become competitive. And even if it is competitive in some markets for say the NSA, that doesn’t make it competitive for general consumer markets. Cryogenic cooling means these things will only work in very special data rooms—so the market is more niche.
The bigger issue though is total cost competitiveness. GPUs are sort of balanced in that the energy cost is about half of the TCO (total cost of ownership). It is extremely unlikely that superconducting chips will be competitive in total cost of computation in the near future. All the various tradeoffs in a superconducting design and the overall newness of the tech imply lower circuit densities. Smaller market implies less research amortization and higher costs. Even if a superconducting chip used 0 energy, it will still be much more expensive and provide less ops/$.
Once we run out of scope for further CPU/GPU improvements over the next decade, then the TCO budget will shift increasingly towards energy, and these types of chips will become increasing viable. So I’d estimate that the probability of impact in the next 5 years is small, but 10 years or more out it’s harder to say. To make a more viable forecast I’d need to read more on this tech and understand more about the costs of cryogenic cooling.
But really roughly—the net effect of this could be to add another leg to moore’s law style growth, at least for server computation.
I guess people aren’t seriously trying this because it’s probably not much harder to go directly to full superconducting computers (i.e., with logic gates made out of superconductors as well) which offers a lot more benefits
It takes energy to maintain cryogenic temperatures, probably much more than the energy that would be saved by eliminating wire resistance. If I understand correctly, the interest in superconducting circuits is mostly in using them to implement quantum computation. Barring room temperature superconductors, there are probably no benefits of using superconducting circuits for classical computation.
Studies indicate the technology, which uses low temperatures in the 4-10 kelvin range to enable information to be transmitted with minimal energy loss, could yield one-petaflop systems that use just 25 kW and 100 petaflop systems that operate at 200 kW, including the cryogenic cooler. Compare this to the current greenest system, the L-CSC supercomputer from the GSI Helmholtz Center, which achieved 5.27 gigaflops-per-watt on the most-recent Green500 list. If scaled linearly to an exaflop supercomputing system, it would consume about 190 megawatts (MW), still quite a bit short of DARPA targets, which range from 20MW to 67MW.
ETA: 100 petaflops per 200 kW equals 500 gigaflops per watt, so it’s estimated to be about 100 times more energy efficient.
What about just replacing the copper wire inside a conventional CMOS chip with a superductor? It took some searching, but I managed to find a paper titled Cryogenically Cooled CMOS which talks about the benefits and feasibility of doing this. Quoting from the relevant section:
So it looks like there’s no fundamental reason why it couldn’t be done, just a matter of finding the right substrate material and solving other engineering problems.
That is the type of tech I was referring to by superconducting circuits as precursor to full reversible. From what I understand, if you chill everything down then you also change resistance in the semiconductor along with all the other properties, so it probably isn’t as easy as just replacing the copper wires.
A room temperature superconductor circuit breakthrough is one of the main wild cards over the next decade or so. Cryogenic cooling is pretty impractical for mainstream computing.
Yeah, its just a question of timetables. If it’s decades away, we have a longer period of stalled moore’s law during which AGI will slowly surpass the brain, rather than rapidly.
From the sources I’ve read, there aren’t any major issues running CMOS at 77 K, you only run into problems at lower temperatures, less than 40 K. I guess people aren’t seriously trying this because it’s probably not much harder to go directly to full superconducting computers (i.e., with logic gates made out of superconductors as well) which offers a lot more benefits. Here is an article about a major IARPA project pursuing that. It doesn’t seem safe to assume that we’ll get AGI before we get superconducting computers. Do you disagree, if so can you explain why?
There was similar interest in superconducting chips about a decade ago which was pretty much the same story—DARPA/IARPA spearheading research, major customer would be US intelligence.
The 500 gigaflops per watt figure is about 100 times more computation/watt than on a current GPU—which is useful because it shows that about 99% of GPU energy cost is interconnect/wiring.
In terms of viability and impact, it is still uncertain how much funding superconducting circuits will require to become competitive. And even if it is competitive in some markets for say the NSA, that doesn’t make it competitive for general consumer markets. Cryogenic cooling means these things will only work in very special data rooms—so the market is more niche.
The bigger issue though is total cost competitiveness. GPUs are sort of balanced in that the energy cost is about half of the TCO (total cost of ownership). It is extremely unlikely that superconducting chips will be competitive in total cost of computation in the near future. All the various tradeoffs in a superconducting design and the overall newness of the tech imply lower circuit densities. Smaller market implies less research amortization and higher costs. Even if a superconducting chip used 0 energy, it will still be much more expensive and provide less ops/$.
Once we run out of scope for further CPU/GPU improvements over the next decade, then the TCO budget will shift increasingly towards energy, and these types of chips will become increasing viable. So I’d estimate that the probability of impact in the next 5 years is small, but 10 years or more out it’s harder to say. To make a more viable forecast I’d need to read more on this tech and understand more about the costs of cryogenic cooling.
But really roughly—the net effect of this could be to add another leg to moore’s law style growth, at least for server computation.
It takes energy to maintain cryogenic temperatures, probably much more than the energy that would be saved by eliminating wire resistance. If I understand correctly, the interest in superconducting circuits is mostly in using them to implement quantum computation.
Barring room temperature superconductors, there are probably no benefits of using superconducting circuits for classical computation.
From the article I linked to:
ETA: 100 petaflops per 200 kW equals 500 gigaflops per watt, so it’s estimated to be about 100 times more energy efficient.
Ok, I guess it depends on how big your computer is, due to the square-cube law. Bigger computers would be at an advantage.