Google imagines out of this world AI - running on orbital datacenters

Google on Tuesday announced a new moonshot – launching constellations of solar-powered satellites packed to the gills with its home-grown tensor processing units (TPUs) to form orbital AI datacenters. "In the future, space may be the best place to scale AI compute," Google executives wrote in a blog post, in which they explain …

[IGotOut]

How to make line go up...

...take a ton of drugs, talk utter bollocks for hours....profit.

[Flocke Kroes]

Calling greater fools

First rule of buying software: do not buy this version for features promised for the next version. (The features cannot be in version N+1 either or you wouldn't need to buy N+2.)

This version of the software does not make a profit. The current investors are at the bottom of a pit with no way to climb out by themselves. They need other people to fill the pit with money for them. Evidently an astronomical quantity of money is required so they their spiel has to be bigger than planet Earth.

[Pulled Tea]

…er, all of this on EARTH ORBIT?

In other words, these compute constellations will need to be quite dense. In one simulation, Google suggests a cluster of 81 satellites that fly between 100-200 meters from one another in an arrangement two kilometers across and at an altitude of 650 kilometers.

Do you want Kessler syndrome? Because this sounds like how you get Kessler syndrome.

[Flocke Kroes]

Re: …er, all of this on EARTH ORBIT?

There were some fun numbers over at Ars Technica comments section:

nVidia DGX200: 8xB200 GPUs, 14.3kW, 145kg

iROSA (ISS's new solar panels): 14kW, 500kg

ISS radiators: 14kW, 740kg

That is 100kg/kW before we include communications, comms power, comms cooling and the struts to hold it together. For a 1GW data centre we need over 100,000 tons of launch. That would be 1000 Starship V3 launches. Boca Chica is liscenced for 25 Starship launches per year so one data centre would saturate Starbase for over 40 years. Add in the proposed capacity of Florida and the time drops to 7 years.

Kessler isn't the problem. The problem is finding investors dumb enough to fund it. (While proof reading I thought of a good place to start looking: the house of commons).

[Pulled Tea]

Re: …er, all of this on EARTH ORBIT?

I mean, yeah! It's expensive, but also…

100,000 tons of potential shrapnel…

[Flocke Kroes]

Re: 100,000 tons of potential shrapnel

That too. Starlinks are designed to burn up completely and even that is a problem on a similar scale to metals dumped into the upper atmosphere by natural micro meteors. We might actually prefer having chunks of data centre splat into the ground compared to depleting the ozone layer.

Turning it around, the ISS has to dodge the occasional dead satellite on its way down. Now imagine a swarm of super ISSs having to scatter to avoid something big.

[Philo T Farnsworth]

Re: 100,000 tons of potential shrapnel

And when those Starlinks and assorted pieces of planned obsolescence come back down, they will, according to one study¹ at Technical University of Braunschweig reported in the web publication Universe Today, fill the stratosphere with "material from satellites that burn up in the atmosphere, especially transition metals [which] could have unforeseen consequences on atmospheric chemistry."²

Universe Today goes on to note that transition metals are "well known for their catalytic activity. Their increased presence in the atmosphere could catalyze reactions that could have devastating impacts on atmospheric chemistry at large. These elements include several that are fundamental to spacecraft construction such as copper (wirings / PCB traces) and titanium (structural supports)."

Further,

"Satellites in [Starlink and Kuiper] constellations are intentionally designed to 'demise' themselves after a few years by burning up in the atmosphere. While this might solve the problem of a bunch of derelict junk that could potentially turn into hazardous projectiles (via Kessler syndrome), it creates its own problem of potentially messing up the atmosphere's chemistry. [...] [There are] different options for dealing with space junk - almost all of them could exacerbate the transition metals problem."

If it ain't one thing, it's another.

__________________

¹ Schulz, et al, Space waste: An update of the anthropogenic matter injection into Earth atmosphere (preprint)

¹ Universe Today: We're Putting Lots Of Transition Metals Into The Stratosphere. That's Not Good.

[tony72]

Re: …er, all of this on EARTH ORBIT?

A healthy dose of scepticism is indeed warranted here. However powering datacentres on Earth is becoming a huge issue, and the availability of "free power" in orbit makes it something that I don't completely dismiss out of hand. Probably you're not just going to take existing server equipment as-is and punt it into orbit, that won't work - they aren't built to be mass-efficient or particularly energy-efficient. You're going to select or design components with energy- and mass-efficiency as a top priority, and your physical form factor could be very different. An obvious option would be to base them on something like a Starlink satellite - a Starlink V2 Mini is estimated to cost around $800,000 dollars to manufacture, and generate around 73.4kW with its solar panels, so that could be your envelope - how much and what type of compute can you achieve with 70kW and 800kg of mass? Launch costs per satellite if you made them exactly the size and mass of Starlinks would be a further $680,000 per satellite on Falcon 9 to LEO, that should come down once Starship is in the game, but that'd have to be factored in. It doesn't seem like an entirely crazy idea to me though, I can totally see Falcon 9s blasting off and deploying an entire orbiting datacentre's worth of flat-packed "computellites" in one go. Maybe there are niche workloads that would benefit particularly from being done in space.

[Doctor Syntax]

Re: …er, all of this on EARTH ORBIT?

"You're going to select or design components with energy- and mass-efficiency as a top priority"

You're also going to select for radiation hardness as a top priority. Is it going to be another of those triangles: energy efficiency, mass efficiency, radiation hardness, choose any two?

[Flocke Kroes]

Re: radiation

Google already tested their TPU with a beam of high energy protons. Scaling for the reduced radiation it would have received in space showed it would have lasted a few years. I got my 100kg/kW by selecting for power efficiency for the electronics and power/mass for electricity and cooling. I did not spare a single thought for price. I also totally ignored connecting the parts together.

How does the heat get from the electronics to the radiator? It is not going through vacuum at a useful rate. Would you like air plus the weight of a pressure vessel to contain it? How about immersion cooling? That would require a pressure vessel to maintain a useful boiling point plus the mass of the liquid. Two phase cooling not going to work without a second hand artificial gravity generator from the United Federation of Planets or the Galactic Empire.

[Flocke Kroes]

Re: …er, all of this on EARTH ORBIT?

Data centres are already optimised for power as it is a hefty chunk of their cost. The other big cost is cooling. A data centre in space cannot dump heat into the atmosphere, a lake or the sea.

Latency is a big problem. Dividing the system into tiny chunks severely limits the problems it can work on and the power required for parts of the data centre to talk to other parts rises dramatically.

Data centres in space are in trouble as a physics problem, a disaster economically and that is before you listen to someone(else) who actually understands how to operate one.

[Doctor Syntax]

Re: …er, all of this on EARTH ORBIT?

"The problem is finding investors dumb enough to fund it."

The existing AI bubble seems to have no problem finding them.

[Anonymous Coward]

Re: …er, all of this on EARTH ORBIT?

It's obvious that none of the people pushing this BS has any idea about cooling stuff in space: "It's cold, cooling is not a problem". Or they are just hoovering money from other idiots and know it.

ISS has two radiator subsystems, approximately 740kg and a bit over 30m^2 each. I'm not sure if 14kW is ISS total or for each subsystem, but still insane: Solar panels and cooling radiator will weight so much that the weight of computers is already irrelevant.

There's no way to reduce either with current technology either, so making 'more efficient computers' is absolutely moot point. Unless the power usage drops to single watts.

I don't see that happening soon either.

[Anonymous Coward]

Re: …er, all of this on EARTH ORBIT?

Or ants, this is how you get ants

[BasicReality]

Cooling these massive data centers has been an issue, I bet putting them in space handles that problem much easier.

[Androgynous Cupboard]

For someone called Basic Reality, you’re sure not grasping it.

Heat. The massive, barely mentioned problem with this ridiculous notion is heat. We are literally boiling away billions of tonnes of water dissipating heat from data centres on earth, so how exactly are we going to do that in space? Black body radiation isn’t going to cut it.

[Filippo]

I assume that's sarcasm?

[steelpillow]

Surely so.

Also looking forward to the formation-swarms sweeping up some of the growing piles of space junk in low Earth orbits.

[The Man Who Fell To Earth]

Skynet

What could possibly go wrong?

[s151669]

Re: Skynet

More like Hactar, I think.

[BartyFartsLast]

Save time

Fire it straight into the sun.

Is it not already bad enough without adding in random bit flips from high energy radiation?

[Robin]

Future Un-Proofing

That's right, let's put all the AI where we can't easily go and switch it off.

[Anonymous Coward]

Heat

How do you dump all the heat in a vacuum?

[MonkeyJuice]

Re: Heat

That's the neat part. You don't.

[Bebu sa Ware]

"powered by a limitless stream of photons from the sun"

Not just any old photon will do - they have to be the right colour,

The rest of the particle zoo streaming out from the Sun might be more problematic especially during a solar flare.

I have the sense these tech megcorps have delegated strategic planning to the "creatives" from their marketing and PR departments. You know, the types you wouldn't trust to open a bottle of beer.

[Flocke Kroes]

Re: delegated strategic planning to their marketing and PR departments

You have it in one. The purpose of the plan is not a viable product. It is to get investment from people who cannot do due diligence.

[frankvw]

What have they been smoking?

I can see so many reasons why this is a terrible idea, I don't even know where to start.

- There's the heavy launch costs, not to mention the technical challenge of getting so many tonnes of Big Iron into space in the first place.

- There's the fact that if Starlink sats have become a troublesome source of RF interference (and they have), what do you think an orbital bit barn is going to do?

- There's the problem with cosmic radiation and solar flares and such, which will necessitate serious shielding, which in turns reduces the available payload capacity.

- There's the orbital debris problem which means that the larger your PV panel array gets (and these will have to be GINORMOUS!) the greater the chance is that something will punch a hole into it sooner rather than later.

- Critical remote server and router maintenance without any physical access to them can be very challenging (I speak from experience) which for kit in orbit is even more difficult.

- AI kit has to be upgraded and extended frequently, which in orbit is simply not an option.

- Et cetera ad nauseam...

Why not go all-out and propose that we simply build AI bitbarns on the moon? I mean, the US is about to land there any day now, using Musks rockets, and well ahead of the Chinese, so how hard can it be, right?

[headrush]

Set the controls for the heart of the sun.

Job done.