[HN Gopher] Superconducting Microprocessors? Turns Out They're U...
       ___________________________________________________________________
        
       Superconducting Microprocessors? Turns Out They're Ultra-Efficient
       (2021)
        
       Author : actinium226
       Score  : 147 points
       Date   : 2024-07-31 01:00 UTC (22 hours ago)
        
 (HTM) web link (spectrum.ieee.org)
 (TXT) w3m dump (spectrum.ieee.org)
        
       | wizardforhire wrote:
       | Another notch on the belt for a cryocooler boom.
       | 
       | Oh "superconducting is hard", "if only we had high temperature
       | superconductors". All valid, but if we work with what we got and
       | disrupt cyrocoolers make them a commodity like magnetrons all
       | those laments become moot.
       | 
       | Prove me wrong.
        
         | Terr_ wrote:
         | So someone else needs to prove a negative, that there exists no
         | possible technology that will "disrupt" cryocoolers by bringing
         | them to an unspecified price/performance point by an
         | unspecified time in the future to service an unspecified
         | computing use-case?
         | 
         | That seems more than a little unfair. :p
        
           | JumpCrisscross wrote:
           | > _someone else needs to prove a negative, that there exists
           | no possible technology that will "disrupt" cryocoolers_
           | 
           | Look at the Wikipedia references for crycoolers [1]. Note the
           | dates and volume. Now look at room-tempuerature
           | superconductors [2].
           | 
           | 1990 vs 2023. 5 vs 57. OP is arguing that a greater fraction
           | of high-temperature superconducting research dollars might
           | find purchase in improving the cryocooler than we presently
           | spend.
           | 
           | [1] https://en.wikipedia.org/wiki/Cryocooler#References
           | 
           | [2] https://en.wikipedia.org/wiki/Room-
           | temperature_superconducto...
        
             | Terr_ wrote:
             | > Now look at room-tempuerature superconductors [2].
             | 
             | As a tangent: Don't forget to look at pressures too. Some
             | newer superconductors that are near-room-temperature aren't
             | quite as exciting when it turns out that requires _over 1.5
             | million times_ normal atmospheric pressure. ( "Hey, I think
             | I over-tightened the CPU heatsink...")
             | 
             | Ex: https://en.wikipedia.org/wiki/Lanthanum_decahydride
        
             | tliltocatl wrote:
             | The problem is, cryocooler theory is pretty well
             | established and "solved at this point, so there is no
             | reason to expect something completely new phenomena there,
             | just some engineering improvements. Solid-state physics, on
             | the other hand, is just computationally infeasible to
             | "solve", so there is a plenty of possibility to discover
             | something unpredicted.
        
         | XorNot wrote:
         | Also notable because heat rejection on that scale till you
         | enter the superconducting regime is a reinforcing loop: the
         | device no longer makes heat, you just need to keep the
         | environmental heat out.
        
           | jon_richards wrote:
           | And the power requirement to maintain temperature scale by
           | surface area, not volume. Same reason thermal power storage
           | using giant piles of sand is theoretically viable.
        
         | yeknoda wrote:
         | Is the cost of cryocoolong capex or Opex dominated?
        
           | hn72774 wrote:
           | Are you referring to the hardware cost vs operating cost? Or
           | how the units are financed?
        
             | idiotsecant wrote:
             | That's what it means. Some things are expensive to buy,
             | some things are expensive to run, and some things are both.
        
             | tuatoru wrote:
             | capex = capital expenditure; opex = operational
             | expenditure.
             | 
             | There's some theorem about investments that says it doesn't
             | matter how they are financed. A good one is good, and a bad
             | one is bad, whether or not you use debt.
        
               | actinium226 wrote:
               | I'll bet you I can find a way to finance a good deal that
               | turns it into a bad one. The other way around seems
               | harder.
        
               | marcosdumay wrote:
               | > it doesn't matter how they are financed
               | 
               | If you have spherical banks in a vacuum, you can simply
               | follow "capital_opex = capex * interest_rate" and then
               | "profit = revenue - total_opex".
               | 
               | But things tend to not work that way on practice.
        
         | nine_k wrote:
         | Would running a cryocooler spend more energy than a
         | superconducting CPU / GPU would save?
        
           | moralestapia wrote:
           | >The 2.5 GHz prototype uses 80 times less energy than its
           | semiconductor counterpart, even accounting for cooling
           | 
           | It's in the header of the article.
        
             | hypercube33 wrote:
             | The way it reads though is that the chip itself uses less
             | power but still needs to be cooled which takes a lot of
             | energy (traditionally)
        
               | moralestapia wrote:
               | >even accounting for cooling
               | 
               | No.
        
               | Grimblewald wrote:
               | stuff like this makes me wonder what the distribution on
               | human context limits is.
        
               | pjerem wrote:
               | 3
        
         | jasonwatkinspdx wrote:
         | I'm fond of this argument but I would point out that magnetrons
         | are remarkably simple devices once you figure out how the waves
         | are interacting.
        
         | tarikjn wrote:
         | I've been thinking about making cheaper cryocoolers for CO2
         | condensation. I'd love to hear your perspective on other
         | applications :)
        
       | d_silin wrote:
       | A more recent take on this subject:
       | 
       | https://diginomica.com/superconducting-chips-could-pack-data...
        
       | incompatible wrote:
       | : The research group in Japan sought to create a superconductor
       | microprocessor that's adiabatic, meaning that, in principle,
       | energy is not gained or lost from the system during the computing
       | process.
       | 
       | I thought that there was a law of information theory that
       | requires expending energy, I think it's Landauer's principle. It
       | seems to be disputed though.
       | 
       | https://en.wikipedia.org/wiki/Landauer%27s_principle
        
         | bryanlarsen wrote:
         | If the superconducting microprocessor is 80x as efficient as a
         | normal micro, that means that it uses about 100 million times
         | as much energy as Landauer's principle. (A normal micro uses a
         | billion times as much, and Landauer's energy is ~1/5th at
         | superconducting temperature).
        
           | incompatible wrote:
           | Yes, but the quote suggests no energy transfer at all. All
           | you need then is a cold place (somewhere in remote space
           | perhaps, or after the heat death of the Universe) and your
           | computation can continue forever without power. It doesn't
           | seem intuitive that that would be possible.
        
             | reaperman wrote:
             | It is extremely likely that the quote is hyperbolic (wrong
             | / inaccurate). If someone broke Landauers limit, that would
             | be a _massive_ headline in academia. Hell even reasonably
             | approaching Landauer's limit would be huge news.
             | 
             | Most likely this headline is a confusion from how
             | signal/power transmission (rather than calculation/work) is
             | truly lossless in superconductors.
        
             | jakobson14 wrote:
             | Continue forever? no. Be performed once at no cost? maybe.
             | 
             | I'm kinda sceptical that a computation to which the 2nd law
             | is indifferent would occur spontaneously without
             | immediately reversing. The 2nd law is what determines the
             | direction that things typically progress.
        
         | KiranRao0 wrote:
         | Per the wikipedia article: "Modern computers use about a
         | billion times as much energy per operation."
         | 
         | I'm guessing the efficiency gains can be considerable before
         | hitting this limit.
        
           | HPsquared wrote:
           | That's actually fewer zeroes than I was expecting.
        
         | Thorondor wrote:
         | Landauer's limit assumes that computation uses a
         | thermodynamically irreversible process and erases bits of
         | information. This is not necessarily true for all useful
         | computers [0]. So theoretically speaking, it's not _impossible_
         | to create an adiabatic microprocessor. Skepticism is definitely
         | warranted though.
         | 
         | [0] https://en.m.wikipedia.org/wiki/Reversible_computing
        
         | philipswood wrote:
         | Energy is expended when you zero or set a bit.
         | 
         | If you compute reversibly you need use special logic gates to
         | not throw any bits away during the computation, like the
         | Toffoli gate. All your operations need to have the same number
         | of input and output bits and needs to be able to run forwards
         | and backwards. Effectively you set or zero no bits during the
         | entire computation that can't be losslessly reversed.
         | 
         | If you structure your computation this way you can do it
         | adiabatically.
         | 
         | You still however need to expend energy when you set all the
         | bits your program requires for execution when you start a
         | computation.
        
           | m463 wrote:
           | I think maybe we need the idea of a "bit well" to match the
           | "bit bucket".
           | 
           | Then it would be possible to just borrow bits when we set
           | bits.
        
           | oceanplexian wrote:
           | My analogy probably isn't perfect but isn't this how an
           | abacus or a slide rule works? You expend energy to set the
           | values for computation but the act of computing and then
           | reading them expends no energy.
        
             | bgnn wrote:
             | Yeah this is a good analogy
        
           | AnthonyMouse wrote:
           | Wait, so does this imply (ignoring the _time_ requirements)
           | that you could do NP calculations with a feasible amount of
           | _energy_ , because the inputs and outputs are small?
           | 
           | Combine that with something that uses time dilation to make
           | it go fast (from our frame of reference) and you'd be giving
           | even hypothetical quantum computers a silver medal.
        
             | Dylan16807 wrote:
             | If you're going to use time dilation so you can wait for
             | the computer to calculate, it's hard to imagine a setup
             | where the power needs of the computer are more than a
             | rounding error in comparison.
        
               | AnthonyMouse wrote:
               | If you're using time dilation the power needs of the
               | computer (in terms of how much fuel you need) are
               | proportional to the rate time passes inside. If the goal
               | is to get the ratio way up there, the power consumption
               | gets important.
               | 
               | Especially if you're using the velocity-based method and
               | have to accelerate the fuel.
        
               | Dylan16807 wrote:
               | You'd be accelerating yourself, not the computer. If you
               | launch the computer it just takes longer.
        
             | Strilanc wrote:
             | Not unless you can isolate the computer from environmental
             | noise exponentially well. Otherwise you'll need to spend
             | exponential energy on entropy removal / error correction
             | (e.g. keeping the dilution fridge running).
        
             | philipswood wrote:
             | I think a longer runtime for a terminating program will
             | require more state to be kept. Every step needs to keep
             | extra reversibility information.
             | 
             | So I don't think this gives any edge on NP.
             | 
             | I suspect that even if you waited for the universe to cool
             | down a lot by waiting aeons and then performed computations
             | arbitrarily slowly you'd still be limited by your starting
             | energy (maximum bits you can write to start with).
             | 
             | Although maybe using random bits might help somehow?
        
               | wizzwizz4 wrote:
               | The amount of energy it takes to write a bit depends on
               | the temperature. But you'd still be limited by the amount
               | of substrate available.
        
             | stracer wrote:
             | > time dilation to make it go fast (from our frame of
             | reference)
             | 
             | The only way to do that is that we move to a place from
             | which the computer processes appear accelerated, e.g. into
             | a strong gravity well. That isn't very useful, because we
             | do not have such a well nearby, as only very dense
             | hypothetical objects can provide it (e.g. black holes), and
             | it would not be compatible with life to move there.
        
               | pennomi wrote:
               | Doesn't have to be compatible with life, just compatible
               | with computing.
        
               | Tuna-Fish wrote:
               | You misunderstand. The problem with using time dilation
               | is that you can only make the clock go slower, not
               | faster. If you wanted to exploit time dilation to get a
               | computer result faster (in subjective time), you don't
               | send the computer to an exotic locale, you go there
               | _yourself_ , while you leave the computer to do work in
               | normal space.
        
       | boznz wrote:
       | Great, I can make my software even more shitty, and nobody will
       | notice.
        
         | 77pt77 wrote:
         | Convenience always wins against quality.
        
       | lytedev wrote:
       | Pardon my ignorance, but would this then be a semisuperconductor
       | or a supersemiconductor?
        
         | kragen wrote:
         | no, they use jjs
        
       | rapsey wrote:
       | Is there any company trying to bring this to market?
        
         | campers wrote:
         | Imec working on this approach too
         | https://spectrum.ieee.org/superconducting-computer
        
       | BenoitP wrote:
       | This is the future. As we try to go 3D and start stacking dies,
       | the limiting factor of heat is getting even more blocking.
       | 
       | It's not about consuming less electricity. It's about dissipating
       | less of it as heat inside the microprocessor. The future will
       | made of tiny porous cubes that are tall sandwiches of RAM and
       | CPU/GPU/etc
        
         | sigmoid10 wrote:
         | Superconductivity with current materials is _hard._ Allost
         | always because of the cooling. You can 't just fabricate a
         | porous cube of transistors and put it in a liquid helium ice
         | bath. You need to have excess heat flowing out in a stable and
         | predictable way. Otherwise once the center reaches critical
         | temperature, the whole thing will explode. There's a reason why
         | everyone is desperately searching for room temp low pressure sc
         | materials. I doubt we'll see industry applications in large
         | scale computing before that.
        
           | mlhpdx wrote:
           | Totally true, but it's interesting that this technology lends
           | itself to incremental improvements. And those improvements
           | can be driven both from the hardware side and the software
           | side. It seems like a virtuous cycle driving this to economic
           | viability and optimization is possible. I'm no expert,
           | though, so this is strictly my imagination.
        
       | est31 wrote:
       | According to the paper, their demonstration chip MANA has 21k of
       | JJ units, which according to their estimates correspond to around
       | 5k transistors. To compare, a single Nvidia GA100 has 54 * 10^9
       | transistors.
        
         | timschmidt wrote:
         | Some versions of the venerable MOS Technologies 6502 have only
         | 3,218 transistors. The Intel 8080 has somewhere between 4,500
         | and 6,000. 5k transistors is square in the middle of "plenty
         | for a classic 8 bit micro". Enough to run a basic *nix or
         | embedded RTOS.
        
           | actinium226 wrote:
           | There's something I'm missing, they say the prototype hits
           | 2.5 GHz, how is that possible if they only have the
           | equivalent of 5k transistors? Or is clock cycle independent
           | of transistor count?
        
             | g15jv2dp wrote:
             | It depends on what you have your transistors do. You could
             | even have one single transistor that you switch on/off very
             | very quickly. You'd need to find a transistor with sub-ns
             | switching time to reach >1GHz. It's not a very interesting
             | "computation," though.
        
             | johntb86 wrote:
             | Adding more transistors doesn't make your clock got faster,
             | and it doesn't increase the speed of an individual
             | transistor. The reason computers got both more transistors
             | and faster in the past was that the transistors were
             | continually shrinking; for a traditional MOSFET, Dennard
             | scaling means that the smaller the transistor (and
             | therefore the smaller its capacitance and voltage), the
             | faster it switches. This device doesn't use MOSFET
             | technology, so its scaling rules are different.
        
             | mrguyorama wrote:
             | I'm very confused but I mean this in the utmost sincerity:
             | 
             | What made you think transistor count and clock speed were
             | linked? What was your line of thought? How did you think
             | overclocking worked, by dynamically removing transistors
             | from the chip?
        
               | RicoElectrico wrote:
               | GP is not entirely wrong. To achieve any reasonable speed
               | pipelining is necessary. That is, breaking up the
               | critical combinatorial path with registers. Which adds.
               | transistors. Cache also is needed, which also consumes
               | transistor budget.
        
               | actinium226 wrote:
               | That's basically my question. I'm not clear on what all
               | the transistors on a chip are doing and I'm trying to
               | understand what the significance of a chip with only 5k
               | transistors but able to do 2.5GHz would be. I imagine
               | there must be some limitation, I'm just not sure what.
        
               | timschmidt wrote:
               | Most of the things modern microprocessors spend
               | transistors on are clever bargains to allow the CPU to
               | execute a single thread faster. Caches which store
               | instructions and data closer than main memory,
               | translation lookup buffers which store already de-
               | referenced memory locations, pipelining which reduces the
               | amount of work and complexity per stage so each can be
               | clocked faster, SMT to make better use of multiple decode
               | ports and execution units, complex additional
               | instructions like AVX for doing more work in fewer
               | instructions, microcode for disconnecting the underlying
               | architecture from the instruction set allowing
               | significantly more design freedom and implementation of
               | legacy instructions without requiring hardware.
               | 
               | A design as simple as an 8 bit micro implements the
               | instruction set directly in hardware, with minimal
               | pipelining, no caches - just a few registers for holding
               | values currently being worked with. They may implement a
               | few dozens to a little over a hundred instructions vs.
               | thousands in a modern x86. It won't have any fancy
               | integrated peripherals like a graphics controller or NPU,
               | just an interface to memory and a few IO pins. Even a
               | 2.5ghz 8bit micro won't be fast compared to a similarly
               | clocked modern x86. The micro may dispatch 1 instruction
               | per clock or per two or four clocks, whereas the x86
               | might decode 6 or 8 instructions per clock per core and
               | have as many as 20 or 30 instructions in flight at any
               | given time per core. But the 8bit micros are just beyond
               | a threshold of complexity which is recognizably a CPU
               | capable of arbitrary computation upon which you can bolt
               | on anything else you might need.
        
       | bgnn wrote:
       | Interesting. To give some context to the current problem: the
       | classical (CMOS) logic gates energy is expended to charge or
       | discharge the load capacitance of each electrical node during a
       | state change. Charging a capacitor to a certain voltage requires
       | E=CV^2 amount of (Joules) energy (C in Farads, V in volts). From
       | this, half of it is the energy stored at the capacitor, other
       | half is the energy converted to heat on the transistors. This is
       | so far the most efficient way we can build large scale integrated
       | logic, and at best it is 50% efficient. In reality there is also
       | unwanted inefficiencies, making it close to 30% or so. Good to
       | keep in mind: 100% of the energy becomes heat because the charged
       | capacitor needs to be discharged at some point to change its
       | state. This is becoming a huge thermal management nightmare.
        
       | dschuetz wrote:
       | How much energy was needed for the cryostat? I haven't found
       | anything on that in the article.
        
       | kragen wrote:
       | unfortunately the title says 'superconducting' when it should say
       | 'reversible'
        
       | Hugsun wrote:
       | I'm surprised that this wasn't done earlier. An experiment like
       | this seemed like an obvious useful thing to try the first time I
       | heard about superconductors, some decades ago.
       | 
       | Perhaps it is due to the technical difficulty of the experiment.
        
       | quux wrote:
       | To quote Norm MacDonald:
       | 
       | ... according to a report in the medical journal "DUH"
        
       | lacoolj wrote:
       | This will be incredible when the cooling at consumer-level has
       | been figured out. I do wonder what kind of material
       | requirements/availability there will be though. Even things like
       | touchscreens and other rare earth metals are either getting
       | scarce or controlled by one or two countries.
       | 
       | Regardless, exciting news here for all of us
        
         | marcosdumay wrote:
         | Lots of people handle liquid nitrogen in all kinds of different
         | field applications. Cooling things is not that hard.
        
       | cashsterling wrote:
       | I worked on a team (at a Company) that built a Scanning SQUID
       | Microscope (SSM) to image magnetic flux trapping in
       | superconducting circuits. The organization that bought the SSM is
       | working on these kinds superconducting microprocessors. A SSM
       | uses a SQUID (superconducting quantum inference device) as an
       | extremely sensitive magnetic flux sensor; SQUID sensor requires
       | further amplification which is also typically a SQUID-based low-
       | noise amplifier (long story). Same Company (I worked for) built
       | cryostats for quantum computing research and all kinds of other
       | cryogenic needs.
       | 
       | There are a lot of issues with designing, fabricating, operating
       | these sort of circuits at large scale... hence the need for a
       | microscope to study flux trapping and other phenomena of
       | operating circuits. But, overall, I'm optimistic that this
       | technology can work.
       | 
       | A note about energy consumption of this technology. Niobium thin
       | film superconducting circuits have to be cooled to about 4 Kelvin
       | to be 'properly 'operational'. Heat leaks from higher temperature
       | stages into the 4 Kelvin cooling zone via conduction of thermal
       | insulating supports, electrical signal lines, and thermal
       | blackbody radiation. Several kW of power are required to provide
       | 1 Watt of cooling power at 4 Kelvin.
       | 
       | There are also small resistive/impedance losses in electrical
       | signal lines connecting room temperature electronics to the
       | superconducting chip. So... I think calling the microprocessor
       | 'adiabatic' is a little disingenuous. Small amounts of power, in
       | the form of many nano-amp and micro-amp currents are required to
       | operate and interface with the chip... the chip cannot operate
       | without this electrical interface.
       | 
       | In additional, in a test environment where researchers are only
       | running one chip... the overall cryogenic system, electronics,
       | and superconducting chip are wildly energy inefficient compared
       | to current microprocessors. But the "forward looking statement"
       | is that hundreds of microprocessors could be run in one cryostat
       | and the 5kW cooling budget would replace the power draw of 100's
       | of classical microprocessors while also provide higher equivalent
       | FLOPS per process processor. But this "forward looking statement"
       | is NOT true today, as far as I am aware.
        
         | written-beyond wrote:
         | I appreciate your comment it's probably very informative to a
         | lot of people here on HN. I will admit this comment for me,
         | read like something out of r/vxjunkies.
         | 
         | Props to you and your team for building amazing stuff. Squid
         | and Niobium are very entertaining names.
        
       | darknoon wrote:
       | Relevant video about some of the history of superconducting
       | computers: https://www.youtube.com/watch?v=14r2oMsAaE8
        
       ___________________________________________________________________
       (page generated 2024-07-31 23:01 UTC)