Supercomputers — Experimenting on What Cannot Be Experimented On

A third method

Science long had two roads: reasoning it out and trying it out — theory and experiment.

A third appeared: computing it. Put in the rules, work out the result after a very short interval, take that as the new start, and step again. Repeat it enough and a billion years pass. That is simulation.

Wrong rules, wrong answer

There is a trap here. However exact the arithmetic, if the rules put in were wrong the result is wrong.

A convincing picture on a screen looks like fact, but it is only what the rules drew. So this work keeps checking results against real observations: compute what is already known, confirm the answer, and only then move to what is not.

A little further in

A supercomputer is not a very fast computer

The name misleads. No single part of it is extraordinarily fast.

Open it and you find tens of thousands of processors much like the ones in ordinary machines, tied together by very fast connections and each given a share of one problem. Not one thing being quick — many things at once.

Half the work is heat

Tens of thousands of processors running together produce enormous heat. Left alone the machine cooks itself.

So half the room is cooling: water circulated to carry heat away, then chilled again, with power held up so a cut does not stop it. The electricity used would serve a small town.

Which joins the energy article. Computing is in the end turning electricity into heat, and that heat has to go somewhere.

The buildings going up everywhere

Data centres are much in the news. They are the same object: a building full of processors being kept cool.

What differs is the purpose. A supercomputer splits one large problem into tens of thousands of pieces. A data centre handles a great many separate requests. One is a team sharing a heavy load; the other is a bank with tens of thousands of counters.

Which brings the same two problems along: power and heat. A large data centre draws what a small city draws, and nearly all of it comes back out as heat. Carrying that heat away takes water, and not a little.

So where to build becomes the question. Cheap electricity and cool air are wanted — hence sites in northern countries, next to power stations, drawing cold seawater, and experiments with sinking them into the sea altogether.

But the opposite pull exists too. Too far from people and the signals take time to travel. The wish to be near users and the wish to be somewhere cold are permanently at odds, which is why data centres crowd around capital regions.

Every search, every question put to an AI, spends electricity in a building somewhere and produces heat there. What looks weightless on a screen is physical somewhere else.

Could that heat generate power?

Everyone has the thought at some point. Power stations boil water and turn a turbine with the steam. Data centres throw off enormous heat. Why not boil water with it and get the electricity back?

A good question, and the answer is mostly no — for a reason this series keeps returning to.

Steam in a power station is hundreds of degrees, far above the outside air. Heat leaving a data centre is thirty or forty degrees if air-cooled, around sixty with liquid cooling. Lukewarm.

Lukewarm water does not boil at all. Fluids that boil at lower temperatures can be used to force the issue, but the return is a few per cent, and the equipment and space needed rarely repay it.

So heat is used as heat

That does not mean throwing it away. Give up on turning it back into electricity, and it becomes useful at once.

Lukewarm water is useless for generation and ideal for heating: warming a room in winter does not require hundreds of degrees. So in northern Europe data centre heat is piped into district heating, or used to warm greenhouses and swimming pools.

The solar and hydrogen article said energy is never made and only changes form. One line can be added: quality falls at every change, and does not come back. What actually disappears while we say we are using energy is not the quantity but the quality.

Then why not build one in space?

Another thought everyone has. Space is around 270 degrees below zero — put a data centre up there and cooling solves itself.

This is where intuition inverts exactly. Getting rid of heat is far harder in space than on Earth.

So heat is a standing problem for spacecraft. Many of the wing-like panels on a space station are not solar cells but radiators. Add unshaded sunlight, no repair crew, and the travel time of signals.

Plans for orbital data centres are nevertheless real — but not because it is cold. It is because sunlight is available around the clock with no atmosphere in the way. If the reason is cooling, the reasoning is wrong. If the reason is power, it is worth discussing.

And a disused mine?

That one is rather plausible. Korea has many closed mines, and data centres already sit inside old mines and bunkers in Europe and America.

The virtue underground is steadiness. The surface swings forty degrees between summer and winter; a little way down the temperature barely moves from the mid-teens all year. It is why caves are cool in summer and mild in winter, and it cuts the cooling load.

Water standing in the workings can serve as coolant, thick rock keeps out storms, fire and intruders, and jobs in a former mining district are no small argument.

It is not simple, though.

1. Damp. Underground is generally humid, and water condenses on machines.
2. Air has to be moved in and out continuously. People and machines both breathe.
3. The ground must be checked for collapse. A disused mine is an abandoned hole.
4. Heavy electrical supply has to be run up into the hills — usually the largest cost.
5. Distance from population means longer signal times.

So the answer depends on the job. Poor for anything that must be instant; good for storing archives or running computation through the night. Less a best site than a question of what work you send there.

As an aside: keep cooling and you may wonder how far down it goes. The floor is absolute zero — though even there nothing quite stops. Matter simply sits in its lowest possible state, and a trembling remains even in that.

What it is used for

1. Tomorrow's weather, next season's, and the climate a century out.
2. Sifting tens of thousands of candidate molecules inside the computer before any real experiment.
3. Watching the instant of a car crash without wrecking a car.
4. Compressing a billion years of galaxy and star formation into a few days.
5. Training artificial intelligence — the fastest-growing share of all.

The institute does more than compute. It runs the high-speed research network and gathers papers and research data. Who keeps observations and results, and how, is a steadily growing problem in modern science.

What you would study to do this

This is not a building of computer scientists.

1. Someone who knows what to compute — a physicist, chemist, meteorologist, biologist in their own field.
2. Someone who turns that into a form a computer can solve — numerical analysis and applied mathematics.
3. Someone who makes it run across tens of thousands of pieces — computing and parallel processing.
4. Someone who cools the machine and feeds it power — mechanical and electrical plant.
5. Someone who keeps it alive through the night — facility operations.

The first line matters. Most people using a supercomputer did not study computing; they are meteorologists or chemists who also learned to compute. Computational science belongs to no single department and touches all of them.

The fourth and fifth lines are real jobs too. The fastest machine in the world stops if a pipe leaks.

The work is done at the Korea Institute of Science and Technology Information in Yuseong-gu, Daejeon.

The question that remainsIf computing stands in for what cannot be tried — what tells us the computation was right?