Superconducting Microprocessors? Turns Out They're Ultra-Efficient (2021) | Hacker News Reader

Superconducting Microprocessors? Turns Out They're Ultra-Efficient (2021)

spectrum.ieee.org

169 pointsactinium22693 comments

: 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

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).

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.

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.

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.

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.

That's actually fewer zeroes than I was expecting.

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

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.

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.

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.

Yeah this is a good analogy

That's a bad analogy because moving stuff in abacus is the computation itself.

Suppose you put the abacus sideways, such that setting a bead requires working against gravity, but then returning the bead to the original position gives you that energy back?

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.

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.

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.

You'd be accelerating yourself, not the computer. If you launch the computer it just takes longer.

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).

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?

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.

> 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.

Doesn’t have to be compatible with life, just compatible with computing.

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.

TIL that time dilation can only really be used to make something go slower than normal space, not faster.

Sci-fi authors owe some apologies. I'm looking at you, Star Trek.

Time dilation means the object in the gravity well is experiencing time at a slower rate than us humans outside. That is the opposite condition we want: us slow, computer fast.

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.

4 bit CPU, about 10,000 gates, 5 GHz. So just a proof of concept at this point.

Cyrogenic computing has been too far outside the mainstream. The mainstream technology improved faster than the cyrogenic stuff.

NSA put large amounts of money into cyrogenic computing, from the 1960s on. "I want a thousand-megacycle computer. I'll get you the money!" - an NSA director. It never really worked out, although at one point some special purpose device, probably a key tester, was actually built. The first round of that cyrogenic technology used cyrotrons. Cyrotrons were fast, but, being magnetic devices, not small enough. The second round used Josephson junctions. NSA finally gave up on that around the time ordinary CMOS passed 1GHz. Lately there's been some interest again.[1]

[1] https://spectrum.ieee.org/will-the-nsa-finally-build-its-sup...

That's the way technology evolves, some bets pay off, while others don't.

Actually, it evolves in a random walk. Old ideas can become more practical because while we abandoned an approach, we got better at things that would make it more feasible another time around (eg VR today looks nothing like the VR of my childhood due to advances across the board in computing power, computer language expressivity, game dev tools, IMUs, IMU algorithms development, etc etc etc). That’s a case of a path opening up due to investment over time that couldn’t be accomplished by a bet.

Other times bets are insufficient in size - a bet that requires 100B to unlock trillions may seem like a failure at 10B, especially if you didn’t properly estimate the size of a market (or the market isn’t large enough for that yet).

Pardon my ignorance, but would this then be a semisuperconductor or a supersemiconductor?

no, they use jjs

A more recent take on this subject:

https://diginomica.com/superconducting-chips-could-pack-data...

Great, I can make my software even more shitty, and nobody will notice.

Convenience always wins against quality.

Is there a hard limit on ensh*tification? I was just posting about Yelp, and reddit from a cell phone, and the thought crossed my mind to just close mobile browsing forever.

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.

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.

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?

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.

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.

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?

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.

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.

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.

Are barrel shifters automatically zero sum? How often do GPUs do renormalization? It's the switching and the location information.

Did we have this discussion 50 years ago? It is a set of brilliant ideas. I read about jjs when I was small, and later re-read about why they never panned out significantly.

CPUs are like fancy tally clickers with an auto-clicker attached. A CPU operating at 2.5 GHz means the clicker is depressed 2.5x10^9 times/second.

What to do with that click speed is technically an entirely different matter.

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

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.

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.

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.

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.

Thanks! Also, I made at least one bad typo. SQUID = Superconducting Quantum Interference Device... not inference.

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

Lots of people handle liquid nitrogen in all kinds of different field applications. Cooling things is not that hard.

Right. But this requires cooling to liquid helium temperatures. That's a separate level of hard and cost compared to liquid nitrogen.

How much energy was needed for the cryostat? I haven't found anything on that in the article.

Relevant video about some of the history of superconducting computers: https://www.youtube.com/watch?v=14r2oMsAaE8

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.

The original superconducting logic schemes were established a long time ago and companies started working on computing circuits as early as the 1960s. IBM spun out its own attempt in the early 1980s when it became apparent that this CMOS thing was going to pan out a lot faster.

Is there any company trying to bring this to market?

Imec working on this approach too https://spectrum.ieee.org/superconducting-computer

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.

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

> 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...

> 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

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.

> cryocooler theory is pretty well established and "solved at this point

What are you quoting? What is "cryocooler theory"?

Thermodynamics. Should have been "solved", missed the closing parenthesis.

The latter aren't being researched for CPUs, which are small things. They're for applications like long-distance power transmission, electric motors, and more. Things which aren't feasible to cryocool.

> long-distance power transmission

IIRC unlikely to change quickly even with higher-temp superconductors, since it would mean splicing in new power-grid segments that transfer direct-current instead of alternating, and then you have losses in conversion too.

We have plenty of long-distance HVDC links, one of the most prominent being https://en.wikipedia.org/wiki/Quebec_%E2%80%93_New_England_T..., which also helped isolate Quebec from the Northeast blackout of 2003 https://en.wikipedia.org/wiki/Northeast_blackout_of_2003

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.

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.

Is the cost of cryocoolong capex or Opex dominated?

Are you referring to the hardware cost vs operating cost? Or how the units are financed?

That's what it means. Some things are expensive to buy, some things are expensive to run, and some things are both.

The terms in that context are fixed cost vs variable cost.

CapEx vs OpEx deals with how the equipment is managed financially. Buy vs lease/rent/IAAS. Good examples here https://centersquaredc.com/blog/choosing-the-best-data-cente...

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.

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.

> 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.

Would running a cryocooler spend more energy than a superconducting CPU / GPU would save?

>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.

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)

>even accounting for cooling

No.

stuff like this makes me wonder what the distribution on human context limits is.

It is very small, but humans tend to internalize knowledge as they read giving essentially infinite but lossy context length. Those posters failed to internalize the message here so they get the wrong knowledge out of the message.

That was both self referential and ignorant of what? Oh. .. forgetting. Or the simulatianiety. Or not. I am stuck in a zero sum post, or not. A comma uses less energy than three dots,

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.

I’ve been thinking about making cheaper cryocoolers for CO2 condensation. I’d love to hear your perspective on other applications :)

To quote Norm MacDonald:

... according to a report in the medical journal "DUH"

unfortunately the title says 'superconducting' when it should say 'reversible'

Reply on news.ycombinator.com