In the discussion of whether and how much RTS (Room Temp Superconductors) help CPUs, two things are getting confused: Using superconductors as interconnects and using some new principle to replace the transistor.
Regarding interconnects:
Elbert wrote:
The problem with making CPU's bigger is limited by how far a given voltage can travel and be read. With superconduction electric can travel much greater distances at much less voltage as almost no voltage is lost.
I'm not sure whether you mean electricity, electrons or electric fields, but this is the 'power transmission' argument and almost entitrely irrelevant. The difficulty to which you refer is the difficulty of transmitting
a signal or more likely
several signals in synchronism across a chip. This difficulty is caused by all interconnects being transmission lines and not plain 'bits of wire' at sufficiently high frequencies, and it applies to all sorts of things from gate-to-gate interconnects all the way to USB cables.
Now, you may have noticed that something like a USB link is pretty slow, but can go quite a way without running out of electricity. This is because the fundametal trade-off is a speed versus distance one and you can get quite a distance if the link can be slow. So what signals have to be fast in your mobo-sized cpu? Well, if you want a CPU that big to have plenty of cores (and, errm, you have a decent interconnect-between-cpus bus, like, oh don't know, say, AMD), the interconnect between cpus doesn't have to have superconductivity to work - look at the 'big iron' Opteron servers, they achieve 'all the way across the mobo' distances today. (And given that someone will get annoyed with me if I mention AMD and I don't say Intel is wonderful too, I have to add 'and I'm sure Intel would have something comparable to HyperTransport, if they needed it for today's CPUs' :? )
There is obviously still the 'Cray love seat' problem when interfacing to main memory (or from the 'cores' to the memory controller), but that goes away for reasonable numbers of cores if you are prepared to spread the memory interfaces around the periphery. Looking at this memory-interface-to-n-cpu problem, you can easily convince yourself that 16 CPUs would be easy-ish, but, by the time you have go to 64, it has become a bit hard. (And I really wouldn't want to try it with 256+ cores, with today's technology, but I can see ways of pushing in that direction with relatively low gate counts per core.) But what is very obvious is that today people aren't building even the relatively easy 16 cores per slice parts, for other reasons.
So, what stops people building mega-big CPUs is the same thing that has always stopped people building mega-big CPUs. Cost of a mega-big piece of flawless silicon. (There is an argument about fault tolerance and being able to tolerate, say, n-1 out of n CPUs functioning on a slice of silicon - this is true, but essentially unaffected by the putative coming of RTS).
Now, given the transmission line nature of all high frequency interconnects, will superconductors enable much better (faster/lower loss) transmission lines to be made? Well, it is not clear that they will, although there might be an improvement. They can make the series resistance component go to zero, but given that the materials that make up the TL also determine the parallel capacitance element, there is a danger that changes in the dielectric constant and loss factor eat up any gains from the removal of the series resistive element.
So there is a danger that superconducting TLs are actually worse than current TLs, due to these other losses and characteristics and it won't be clearer until actual materials are proposed. And, anyway, that only gets the signals to the MOSFETs...they still have to switch, so to get an order of magnitude improvement, you'd still need roughly and order of magnitude improvement in the FETs that are to switch.
And it is the threshold of the FETs which is the prime determinant in setting the operating voltage levels of the technology. Whether you have superconducting interconnects or not, you still have to supply enough voltage to switch the FETs on and off. So superconducting interconnects aren't by themselves going to deliver order of magnitude lower operating voltages.
Elbert writes:
Copper used for traces and gates has been used for a long time as about the best cheap semiconductor.
In case anyone is worried, copper hasn't suddenly become a semi-conductor, it is still a conductor.
The link that you present
(hafnium paper) doesn't allow you to get the paper, but the title of the abstract tells its own story "The Anomalous Behavior of the Dielectric Constant of Hafnium Silicates: A First Principles Study". This is about dielectric properties and not conductors and so is interesting but not directly germane.
Now crosstalk only limits how close traces can be which is the worst cause of the bigger problem of leakage.
Well, there are two types of crosstalk and it is not clear which one you are referring to, but neither cause leakage. True leakage is a static phenomenon and you measure it by disabling switching. Crosstalk is a dynamic phenomenon and only occurs when the dynamic signals are in play. The whole area of signal integrity is a difficult one and you need to be clear about which effect is which.
As for Josephson junctions and quantum computers, this is on the edge of massively weird technology compared to what we have today. JJs have been the next superfast thing for pretty much 40 years. They are still the next superfast thing, in some quarters. The quote (Stuart72):
I seem to recall these things switch 1000-10000 faster than silicon
.
has been around for some time. Trouble is, 40 years ago, silicon switched 1000 times slower than today, so some of the advantage that JJs once would have theoretically had, has been eaten away by the year-on-year progress with more conventional technology.
If you have a superconductor that is has different resistance based upon an electrical field, then you have pretty much the same device and can completely design digital electronics using this material.
While this is true, it is not clear that it will turn out to be an advantage. After all, you still have to make the mag field that turns on and off the superconductor (and that probably changes with temperature) so it could be that the problem suddenly becomes getting enough current to switch the superconductor on and off rather than enough voltage to switch the FET on and off. And if, in order to do that, with the values of magnetic field that are needed, you have to produce something like a micro-induction coil, the inductance of that arrangement will limit the rate of curent rise, and that will slow switching.
And if you contrive a structure that switches states at some very low field threshold, in order to circumvent this problem, then the field produced by the load current is in danger of switching the device, which is not generally what you want.