It has been a week and thanks to
tests by a few Chinese scientists, we can probably say that LK-99 isn't
a room temperature superconductor. The hype has largely died down. LK-99 has vanished
off people's social feeds and science news. I suppose we can still keep on waiting
but the right now it's so over. Oh well, it wouldn't have done anything
for the semiconductor industry anyway. I did some follow-up work since the last video. Even if we hit the jackpot tomorrow and find
a True Blue room temperature superconductor, it will be nice but hardly
a revolution. In this video, some irresponsible thoughts about
superconductors and semiconductors. ## Interconnects The first thing I thought about when I thought about superconductors for
chips were interconnects. In an integrated circuit,
interconnects are wires for transmitting the electrical signals between
the transistors and other circuit elements. These interconnects are usually but not
always made from low-resistance metal, so we call them "metal layers". The wires are surrounded by dielectric
material - "intermetal dielectric layers" that serve as insulators. For a long time,
we used trusty silicon dioxide for this. These layers sit all jumbled about on top of the silicon transistors like how
stratum look like in geology. Generally speaking, an IC has two types
of interconnects - local and global. Local interconnects occupy the chip's
lowest levels and connect local circuit blocks. The wires are very short so
designers are less concerned about line resistance than they are
about other factors like heat. So I don't think a room-temperature
superconductor - which would be quite sensitive to heat - makes sense here at all. Global interconnects are higher up in the
stratum. They connect distant parts of the chip - delivering important signals like synchronizing
clock signals or power to run the transistors. For this reason, they have great
size. They look thick. Solid. Tight. The distance they have to
span also means that they must be made from low-resistance
metals like aluminum or copper. ## Superconductive Interconnects? I was very curious about this. In today's most advanced AI accelerators - like
the Nvidia H100s we use for massive models like ChatGPT - some 80-90% of the power budget is spent
on simply transferring data around rather than actual computation. Will superconducting
interconnects simply wipe that out? So I asked MIT PhD Alex Sludds for a few thoughts.
He is a photonics researcher with a background in semiconductor devices and interconnects. And
of course, he is a good friend of the channel. I did a video about his work in neural
networks and photonics a while ago. Alex said: > At short lengths of wire,
having a superconducting wire will help significantly decrease the energy
consumption to transmit a zero or a one. When he mentions "short", he means as
in the context of wires on the chip or between multiple chips like in a chiplet scenario. > However, this energy will not become free, since wires will still have a
capacitance and an inductance that stores in the electrical/magnetic field of
the current flowing through the superconductor. What I think Alex is saying - if I get it wrong
it's because I'm too dumb to understand his words - is that even if the interconnects suddenly
possess superconducting superpowers, its electric current will still interact with its surrounding
environment in a way to cause it to leak power. When you push a current through a wire, that wire
creates an electric field around itself. Some of the current's energy gets stored in that electric
field, causing loss. This is "capacitance". The wire will also produce a magnetic field, and that magnetic field will store some of the
current's energy too. This is "inductance". So in terms of energy consumption, there’s going
to be some improvement. That will especially be the case for global interconnects for transmitting
power rather than communication signals. But the power consumption gains might not be as
dramatic as I might have initially thought. ## Speed What about computing speed? A substantial portion of the computer's cycle
time is spent waiting for data to arrive. So, if the interconnects can carry the data
around faster then that's great. But there are fundamental limits
on how fast we can make this. The absolute ceiling in signal speed -
the fastest that anything can possibly travel through any medium - is the
speed of light. And unfortunately, signals cannot travel at light's full
speed because the wire is not a vacuum. Additionally, we need to consider a
form of interconnect delay known as "Resistance-Capacitance" Delay or RC
Delay. Like as I mentioned earlier, the interconnects store charge. These delay the
electric signal's transmission to the receiver. Capacitance depends on the dielectric material
surrounding the wire. That material has a dielectric constant, called k-value. It has
little to do with superconductors. So even if we cut the line resistance to zero, we still
have to deal with the capacitance part of the delay equation. I wonder if the speed gains
will be as significant as I initially thought. ## Current Density There is one last thing about the
interconnect thing that I want to mention. Remember in the first superconducting
video when Kamerlingh-Onnes - the discoverer of superconductivity - tried testing
superconducting lead wire to generate electricity? He discovered that superconductivity requires
not only a low enough temperature but also a low enough magnetic field and
a low enough current density. After the discovery of High Temperature
Cuprate superconductors, people investigated using them for interconnects
in superconducting integrated circuits. Such superconducting interconnects - when
matched with the right circuits - can indeed get very fast. The concern
however is the current density ... Which is defined as the electric current divided
by the superconductor's cross-sectional area. As it turns out, the cuprates' critical current
density limits are not very high. So in order to run integrated circuit operations and stay
under those limits, the interconnects have to be very wide. Perhaps even a centimeter wide,
which for most digital circuits is unfeasible. This is why the cuprates never made headway
into the integrated circuit arena. LK-99 or any room temperature superconductor
may suffer the same issues. If so, then it is just a scientific curiosity
and copper retains its dominance. ## Manufacturability Even if the new superconductor somehow
miraculously fits into the current IC "box", we have the manufacturability concerns. I asked Dylan Patel - chief Analyst at
SemiAnalysis - to say a few thoughts. He said: > While LK-99 seems to show some promise as
either an extremely low resistivity material or in fact even potentially a superconductor the
manufacturability of it is still in question. > While it does seem to be a fairly
simple process for a low purity, low volume amount, scaling this at
high volume at high purity with less than 1 part per billion defect is
going to be a tremendous challenge Dylan has a very good point. TSMC has
said that a single integrated circuit can use over 100 kilometers of copper wiring.
They produce billions of chips every month. To me, this is the biggest issue. Any room
temperature superconductor will be a complex alloy and those tend to not scale. How can we
integrate it into the existing process flow? It took 15 years for IBM to crack the complicated materials engineering issues
surrounding copper interconnects. And that's copper. We have been
playing around with copper for 10,000 years. Any room temperature superconductor
will have been known for far shorter. ## Josephson Effect In 1962, the British physicist Brian
Josephson - then a grad student just 22 years old at Cambridge - wrote a paper predicting
a phenomena that we now call the Josephson Effect. The effect occurs when you put two superconductors
next to one another with a certain thin insulating barrier in between them. Very thin,
as in just a few atomic layers thick. The Josephson Effect produces what is called
a "supercurrent", where a direct current flows through from one side of the device to the other
despite the physical presence of an insulator. The supercurrent flows without resistance and
with little energy loss. This is in contrast with traditional transistor gates, which generate
heat as they switch. That heat must be dissipated. The Josephson Effect is one of the few examples of quantum mechanics - quantum tunneling -
that we can observe. So that is pretty cool. Josephson later won a Nobel Prize for this. Man,
they just handing out Nobels for superconductors! No really, they are. Superconductivity
research has won a tenth of all the physics Nobel prizes awarded
since its discovery in 1911. ## Josephson-junction Electronics A year after Josephson's thesis came out, P.W. Anderson and J. Rowell of Bell
Labs produced an experimental circuit based on this theory, creating
the Josephson Junction Switch. There are a variety of implementations.
The most popular uses niobium as the superconductor and aluminum oxide as the
insulator, creating a sandwich structure. They can switch on or off in just
a few picoseconds, or a trillionth of a second. We take advantage of this by
encoding the data into very rapid pulses, allowing processing speeds
of 100 Gigahertz or higher. I don't think anyone disagrees that
these can get very fast. But there are major issues. Josephson Junctions
require big compromises and suffer shortcomings in certain aspects.
Let me go through a few of them. ## Memory & Complexity IBM worked on producing Josephson
Junction-based computers for two decades, eventually closing their efforts in 1983. One of the major reasons why they finally ended
the project was that they struggled with producing a proper superconducting memory circuit.
Memory is basic and essential to a computer. The largest Josephson-only memory unit
ever produced was a 4-kilobit unit produced in 1990 by NEC. The reason
the memory issue exists is that it is extremely difficult to produce
these junctions in large densities. Similarly, this is why the most successful
Josephson-based systems have been analog systems and the like. Analog systems
handle real world signals like power or sound or the like. They have only
a few components so it’s easier for superconductor-based integrated
circuits to compete with them. Sensors too, Alex pointed out that we can
use superconducting circuits to produce some of the most sensitive photon detectors -
superconducting nanowire single photon detectors. Digital systems on the other hand are built
with millions and billions of consistent transistors. It will be far harder to produce
a competitive superconducting digital system. Something that might make the most sense would be
like a semiconductor-superconductor hybrid module, like AMD's chiplets. This is where we combine semiconductor-based memory systems with
superconductor-based processing systems. ## Freezing Concerns So even if we get a True Blue, Real
Deal room temperature superconductor. And it fits in the IC critical limits box. And
it can be manufactured in sufficient purities. And we figure out how to make dense bunches of
Josephson junctions with it. And maybe we do that hybrid semiconductor-superconductor modules
we talked about to overcome the memory problem ... it still doesn't mean that we get to throw off
the freezing equipment like Prometheus's chains. As a superconductor approaches its
upper limit transition temperature, the number of Cooper pairs in
the superconductor declines. If you recall, Cooper pairs are essential
to making superconductivity work. The less Cooper pairs we have, the
more unstable the circuits will be in the real world. Rule of thumb
per BCS theory is that we want to keep the system's operating temperature
at half of its transition temperature. That is why existing superconducting
Josephson circuits made from Niobium are kept at 4-5 Kelvin despite Niobium's
transition temperature being 9.3 K. LK-99's claimed transition temperature
is around 400 Kelvin. So its practical working temperature would be around 200 Kelvin. Or around -99.67 degrees Fahrenheit, or -73
degrees Celsius, or 360 degrees Rankine, or -9.055 degrees Wedgwood, or 179 degrees Leiden If we want a true blue room-temperature
superconducting chip without the freezing equipment, we need to find something more
like an oven-temperature superconductor. With a transition temperature of 600
Kelvin or 620 degrees Fahrenheit. Refrigeration remains the single biggest
limiting factor in superconductor-based computers - adding a big energy
usage cost on top of the system. The dream that we all have - the tech-optimist’s
fantasy - is that superconductivity at room temperature is now possible once we get
the room temperature superconductor. I was heartbroken when I
learned that it wasn’t so. ## Conclusion If LK-99 is the real deal, it will
be a historic scientific landmark. It might also do impressive things for the power
transmission and energy industries in the real world. Though I am just saying that off the top
of my head without having done the research. But superconductors aren’t serious candidates for
disrupting existing semiconductor technologies. At the risk of being made an idiot later, I am saying
that I don’t believe that any superconductor-based computer will become commercially
competitive at any time in the future. But hey, the hype was a
blast for about a week or so.
