Google launches first datacenter satellite and research that finds orbiting bit barns can work

Google’s first “Suncatcher” datacenter satellite has made it into space, but the company has cast doubt on whether its orbital ambitions can scale. As The Register reported last week, Project Suncatcher aims to test whether Google’s tensor processing units (TPUs) can survive a launch and the harsh environment of space. The Big G …

[Groo The Wanderer - A Canuck]

Not to mention the @)&@#)%&@) pollution of the night sky with glittering shit-boxes streaking everywhere...

[Anonymous Coward]

I don't like it either.

Much as I dislike the idea of sticking these things in orbit, light pollution isn't as bad as it appears. These will all be in the same orbital plane where they get permanent sunlight, so they'll all be overhead at dusk and dawn, rather than at any random time of night. Just about when astronomers are getting ready or packing up.

[Charlie Clark]

Re: I don't like it either.

You forgot to add… for any place on earth as they're not going to be in geostationary orbit just over one place.

[Groo The Wanderer - A Canuck]

Re: I don't like it either.

I didn't say "light pollution"; whether the satellites are illuminated is completely irrelevant to whether they interfere with the field of view for earth-based observation telescopes. That's an ignorant, short sighted, and very stupid decision to saddle humanity's growing knowledge pool with, and very typical of that damned moron Musk: his only concern when making decisions is whether he thinks he can turn a profit. Fuck the environment, fuck the citizens, fuck the community, fuck the nation, fuck the planet: I made money.

[jake]

And the race ...

... to completely clog LEO is on!

And for what? A completely useless flight of fancy from the kiddies brainwashing the AI investors.

[DS999]

Re: And the race ...

I think the bubble is going to burst before we see any more than a few test launches of orbital datacenters.

[Androgynous Cupboard]

$200/kg

I'm clearly in the wrong business.

Dear Google, I'd be happy to undercut that and will only charge $150/kg to put your computers in a truck and drive them to an existing datacenter.

[that one in the corner]

Re: $200/kg

And in two years (18 months? 6?) when the boards are deemed too old and not cutting-edge enough, charge a modest fee for disposal[1] and flog the RAM to the desperate PC users[2]

[1] to pay for the butane flames distracting away from your trucks - they want to see it go up in fire, just like a real re-entry

[2] and make a killing with your adapters as well

[Irongut]

Re: $200/kg

> flog the RAM to the desperate PC users

HOW? HBM is integrated into the GPU package and will not work in a video card or motherboard.

[Like a badger]

Re: $200/kg

For sale: Former Google GPU unit, ideal for playing CounterStrike at max resolution or running badly conceived large Excel spreadsheets.

The system comprises a 42U chassis that contains 4x8 Nvidia H200 and 4.5 TB of HBM3e VRAM. Ideal for gaming, this compact (205cm high, 60cm wide, 48cm deep) is barely larger than a tall kitchen cupboard, and weighs just under 750kg. Easily installed with widely available 3 phase 415V power, the system consumes a maximum of 80kW, and is ideal for DIY modding to liquid cooling. Be the envy of your gaming mates! Bid now!

[DS999]

Re: $200/kg

They are not suitable for doing any "GPU" tasks because they don't have video output, or even the video driver circuitry.

[that one in the corner]

Re: $200/kg

They could be used for any compute, including calculating the complicated contents of a video buffer. Which can then be shipped to a cheap'n'cheerful GPU-less video card (just like we used to have) for display. No idea if that'll upset gamers but it'll work for modelling etc.

As for not being able to access the RAM from a CPU - ah, there is plenty of compute very closely integrated with that RAM, just use that! Plug it into a cheap'n'cheerful mainboard with a measly little CPU to handle the housekeeping of the peripherals, like USB keyboards and mice, a NIC; nothing fancy, just enough to get the big fat "GPU" working. You can even run an x64 emulator on the "GPU" and fire up Windows: the grunt is there.

In fact, the only problem is the physical: the card interfaces and whether they can be liberated from the "42U chassis" and "80kW" PSU into more manageable units. Or, if they are cheap enough - and we're talking about rescue from landfill - can be sliced into less-power hungry pieces. With a hacksaw.

Whether this could really be done isn't a technical question. The issues to get past are things like not looking good when put into a PC case with glass sides and RGB LEDs; or simply getting at the kit before it is burnt up for the insurance money.

[that one in the corner]

Re: $200/kg

The HBM is integrated into - a thing with lots of compute that could run the PC users' programs.

Or just run a single routine (in parallel) to shuttle the data to and from the boring old CPU and the HBM. We are long past the mainboard memory slots being a parallel bus of address and data lines so such a scheme is just making explicit what goes on inside your PC today.

As for "the DIMM slots cannae take it" - sod 'em. We've only got DIMM-slot-du-jour one the mainboards 'cos that is how they've agreed to package it. Even today's CPUs don't care if the RAM is in a DIMM or not, so long as the signals get through. You could even run an adapter cable - yes, I said cable - from a DIMM slot adapter to a honking great pile'o RAM that happened to be inside a cobbled together monstrosity that resulted from you, a soldering iron, hacksaw and an abandoned DC. Yes, you'll have to slow down the access speed from your CPU. But your next iteration will use some PCBs you have ordered and are waiting on Evri to deliver. And shortly after, someone who really knows what they are doing have put together a board with nice short, balanced, traces and off we go.

The situation being envisaged is the liberation of huge amounts of RAM (and compute, but stick to RAM) from landfill in a time when there simply is no manufacturing capability to build (affordable) DIMM (ooh, look, many DIMM fab lines are already shut down) and won't be for years. So are there any routes to let the individual PC users use the RAM?

If you want to shoot down the idea of repurposing the hardware from load of defunct DCs the only arguments that really fly are not technical but financial and economical (with a side order of ecological and even political): do we want to use that RAM, do we want to use it locally, has the bank that owns the debt and now the DC got the patience to wait or just wants to render it down for the few grand's worth of gold on the thinly-plated connectors?

[jpo234]

Re: $200/kg

Those $150/kg you are offering need to include electricity to power the kit for around 5 years.

[that one in the corner]

Significantly larger and entail much closer formation flight

Perhaps have them in close enough formation that they can use, what are those things called, ah yes, cables.

Cheap and effective rendezvous and docking are techniques that are worth developing and eminently sellable technology. Hmm, wonder if the first one out can trail a cable from Starship that the others drop a loop around and then winch themselves together. Maybe not, cold gas jets it is then. But still best to do it soon after launch, whilst they are still relatively close together. Can still chuck away any units that didn't survive the launch and being pezzed

One, well, space station could have advantages (like, one shiny thing in the sky rather than lots spread out all over). Nice short data paths to start with, and sharing of resources like power and communications back home. Send up a variety of units, like replacement solar cells or compute, as needs change.

No matter how daft we think the idea of space-borne bitbarns is, these guys all seem hell-bent on trying it, so lets at least think of ways we could gain from their nutty actions.

[Stevie]

Re: Significantly larger and entail much closer formation flight

Cables of any meaningful length increase the vulnerability footprint wrt EMR, at least if the network experience from a machine tool workshop floor is any guide.

Here on earth the fabled anti-drone EMP doesn't work well enough to be practical (I checked) but out in space massive EMP can be a way of life on account of someone leaving an unshielded fusion reactor eight anna half light-minutes away.

Your (admittedly very attractive) space station would need to be coated in several feet of water or ice to be safe(ish) and that would make it more massive and expensive to keep aloft in LEO.

But sticking the nasty humming, water-polluting, energy-sucking AI-sheds in space is a laudable goal.

If only we could send some of those building the things there too ...

[that one in the corner]

Re: Significantly larger and entail much closer formation flight

> Cables of any meaningful length increase the vulnerability footprint wrt EMR

Cables? Hmm. Do you mean the cables for hauling the boxes together (which I poopooed already, sad face) or do you mean the interconnects? 'Cos if the latter, then - we already use fibre-optic cable for corruptible data lines and you always filter your power lines... Or did you mean to type "EMP" here as well, not "EMR"? 'Cos that then raises some worrying questions regarding "the network experience from a machine tool workshop floor is any guide"!

> EMP ... ice ...

Um, did you miss the entire part about this station idea just being a "well, if they going to do it then let's see if we can make a good fist of it" and "practicing rendezvous is a useful skill"?

The agglomerated space station idea isn't adding any robustness against EMP. Or against an astronaut with a really big hammer (and a lot of gas in his thrusters to counteract - look, he has a space-hammer and it works, ok?).

[Tron]

So Google plan to kill astronomy

Like we needed any more reasons to hate GAFA.

[Lon24]

Re: So Google plan to kill astronomy

Not at all. It's a techbros business opportunity to shoot your telescope into HEO and take out a Starlink subscription to view. Better check the weight and your credit rating ;-)

[Charlie Clark]

Useful research

Rather than tout bullshit about the possibilities like Musk did just to inflate the shareprice, Google is doing some actual research and points out that "limitless" solar power isn't really enough to make the idea viable.

Commercial pressures back on Earth are already pushing data centres to become more efficient as well as to use the cheapest and most reliable source or power, which means that current notional advantages or "data centres in space!" may not be the same in a couple of years. And this leaves out cooling, maintenance replacement cycles and resilience.

[Jimjam3]

As with most things, maintaining equipment is a necessity. Data centres included. Not sure having it in Space is a smart move.

[Irongut]

Bit Barn?

More like Bit Shoebox.

[Anonymous Coward]

Make them orbit the moon

or preferably Mercury and they they won't pollute the world with their garbage.

I'm sure that these DC's that generate AI Slop will soon become prime terrorist targets if they already aren't. Watching those drones destroy the Russian version of Amazon was enlightening.

I wonder when we'll see DC's with anti-drone defences?

This is a mad, mad, mad world we live in.

Oh, and AI is not Superior Intelligence unless you are a numbskull named Donald 'I'm sorry I haven't a clue' Trump.

[Rich 2]

Why?

Why would you want to put your data centre in orbit?

It’s a hideously harsh environment. It’s very expensive. Very difficult to maintain. Cooling is a significant issue. And even if you get it to work you then have to get your data to and from it. Which requires more, very expensive, kit on the ground.

What do you possibly gain?

[druck]

Re: Why?

Less issues with planning permission?

[jake]

Re: Why?

What, you really think that lot gives a flying fuck about planning permission?

[ecofeco]

Re: Why?

I had to scroll this far down.

What is the effing point of doing this?

[Fruit and Nutcase]

Recycling?

No chance of recovering whatever from the kit once its EOL or faulty, its discarded in space or disposed of by burning it up in reentry.

I'm going to get a thicker tin foil hat

[Nik_S]

How do you upgrade these things?

They skip over a big problem; the inevitable advances in technology. With a data center on earth you can in and replace racks full of servers with new equipment relatively easily. How would you do that for satellites in orbit? This is particularly worrying for two reasons:

1. We are still at the very early stage of AI-acclerator technology so newer better hardware will keep coming at a significant pace, particularly as some of this AI capacity will inevitably be used to design better systems

2. We can't just junk these satellites, burning them up in the atmosphere isn't a great solution, and pushing them into ever higher orbits is very expensive