Google's Next Moonshot Is Putting TPUs In Space With 'Project Suncatcher' (9to5google.com) 48
Google's new "Project Suncatcher" aims to launch Tensor Processing Units (TPUs) into space, creating a solar-powered, satellite-based AI network
capable of scaling machine learning beyond Earth's limits. Google says a "solar panel can be up to 8 times more productive than on earth" for near-continuous power using a "dawn-dusk sun-synchronous low earth orbit" that reduces the need for batteries and other power generation. 9to5Google reports: These satellites would connect via free-space optical links, with large-scale ML workloads "distributing tasks across numerous accelerators with high-bandwidth, low-latency connections." To match data centers on Earth, the connection between satellites would have to be tens of terabits per second, and they'd have to fly in "very close formation (kilometers or less)."
Google has already conducted radiation testing on TPUs (Trillium, v6e), with "promising" results: "While the High Bandwidth Memory (HBM) subsystems were the most sensitive component, they only began showing irregularities after a cumulative dose of 2 krad(Si) -- nearly three times the expected (shielded) five year mission dose of 750 rad(Si). No hard failures were attributable to TID up to the maximum tested dose of 15 krad(Si) on a single chip, indicating that Trillium TPUs are surprisingly radiation-hard for space applications."
Finally, Google believes that launch costs will "fall to less than $200/kg by the mid-2030s." At that point, the "cost of launching and operating a space-based data center could become roughly comparable to the reported energy costs of an equivalent terrestrial data center on a per-kilowatt/year basis."
Google has already conducted radiation testing on TPUs (Trillium, v6e), with "promising" results: "While the High Bandwidth Memory (HBM) subsystems were the most sensitive component, they only began showing irregularities after a cumulative dose of 2 krad(Si) -- nearly three times the expected (shielded) five year mission dose of 750 rad(Si). No hard failures were attributable to TID up to the maximum tested dose of 15 krad(Si) on a single chip, indicating that Trillium TPUs are surprisingly radiation-hard for space applications."
Finally, Google believes that launch costs will "fall to less than $200/kg by the mid-2030s." At that point, the "cost of launching and operating a space-based data center could become roughly comparable to the reported energy costs of an equivalent terrestrial data center on a per-kilowatt/year basis."
yikes (Score:1)
Do you want Skynet? 'Cause that's how you get Skynet.
Re: yikes (Score:1)
Re: (Score:2)
They are saying it does make economic sense.
From the summary: Google believes that launch costs will "fall to less than $200/kg by the mid-2030s." At that point, the "cost of launching and operating a space-based data center could become roughly comparable to the reported energy costs of an equivalent terrestrial data center on a per-kilowatt/year basis."
If that's true then space based AI farms are inevitable.
Re: (Score:2)
Sure, you've got cheap power then, but how the frack are they planning on cooling the TPUs? Heat disposal is a real issue for spacecraft, the first thing the Space Shuttle would do is open the bay doors to expose the radiators to space so the habitat area didn't overheat.
Re: (Score:2)
It's not happening. Good god folks, have you heard of any means of launching Satellites into space yet that doesn't require rocket fuel? No? Then it's not going to fall to $200/kg.
The closest we are getting to that is designing a quadcopter that can reach a high enough altitude self-powered before it switches to solar powered ion-drive, which still requires a propellant. You gonna launch a satellite that has no fuel? Good luck.
Unless the material science improves enough to build a space elevator, we're stuc
Re: (Score:2)
They are saying it does make economic sense.
They are wrong.
From the summary: Google believes that launch costs will "fall to less than $200/kg by the mid-2030s." At that point, the "cost of launching and operating a space-based data center could become roughly comparable to the reported energy costs of an equivalent terrestrial data center on a per-kilowatt/year basis."
If that's true then space based AI farms are inevitable.
Except, go weigh an AI data center, multiply that weight by $200/kg, that's your unique cost a land-based AI data center won't have. Now, divide that unique cost by 5 years, the anticipated useful life of this flying data center, that's number is how much cheaper the electricity will need to be in space to make this viable. NOTE: I haven't included the cost of the added weight need to be lifted for the solar panel array to power this flying shipping container data center, let alone the cost of the solar panel array necessary.
Put simply, they promise that a flying data center with its own inexhaustible power source can be operated in space more economically than a similar, self-powered data center operating on the ground, despite the (very optimistic) lift cost of $200/kg... That's a fantasy.
Once you get the flying data center components up in orbit, will the thing be pre-built like the Hubble telescope, or will it need to be assembled, like the ISS from multiple maned launches and assembled in space? If it requires one (or more) manned launches to assemble, that adds to the project lift weight in a not so insignificant amount.
Re: (Score:2)
it may make more sense to build it on the moon and then run a cat5 cable to earth.
What if... (Score:3)
What if we just got all the AI tech bros and put them on these rockets so they can live in space in harmony with their perfect AI computer farms.
Re: (Score:2)
I've been saying this all along - why wait? They aren't getting any younger. Elon and Petey should gather up their most trusted minions and jet right off, the sooner the better.
Stop teasing, just fucking go Galt already. I promise to pretend to miss you if it'll help.
Heat ? (Score:4, Interesting)
How are they planning to dissipate the heat from all this computing? It's a significant problem in space especially if your orbit is in the sun all the time.
Re: (Score:2)
They will conveniently find a way to convert the heat into kinetic energy to control its orbit, and recycle it back into electricity to run the hardware and lasers.
Oh yeah, there WILL be lasers.
Kind of like what they should have been doing with the waste heat from all of those terrestrial server farms; converting it to kinetic or electrical energy.
Re:Heat ? (Score:4, Informative)
They will conveniently find a way to convert the heat into kinetic energy to control its orbit, and recycle it back into electricity to run the hardware and lasers.
The laws of thermodynamics guarantee that you must have waste heat if you are producing useful work. You can't "recycle" waste heat into useful energy (if you could, it wouldn't be waste heat), except by finding a colder place to reject it to.
Re: (Score:2)
So you're saying that there is no way to take the heat and run it through some sort of zero gravity steam turbine that converts the heat into electricity that can then be used by the system?
Re: (Score:3)
So you're saying that there is no way to take the heat and run it through some sort of zero gravity steam turbine that converts the heat into electricity that can then be used by the system?
That would be what is called a "bottoming cycle"-- you take the waste heat and use it. But you still have to reject the waste heat, and you have to reject it at a colder temperature than the (hot side) temperature of the bottoming cycle.
So, for example, if your system is capable of rejecting waste heat at say 350 Kelvin, you could run your electronics at 400 Kelvin. Now if you reject your waste heat at 350 Kelvin, you can use the fifty-degree temperature difference to generate power.
That's the trade off: if
Re: (Score:2)
So the Seebeck effect can't be used recursively while the orbital passes through the shadow of the Earth?
Re: (Score:3)
The article specifically says that Google intend to launch this into a sun-synchronous orbit meaning there are few or no dark periods during which heat can easily be radiated away.
Re: (Score:2)
Darn :(
Re: (Score:2)
they can run power cable to mirrors on sunny and datacenters in shadow. without taking any physics into account (because thats work)
Re: (Score:3)
So the Seebeck effect can't be used recursively while the orbital passes through the shadow of the Earth?
Seebeck effect (aka thermoelectric generation) requires a hot side and a cold side. So, to generate power from this, you have to run the hot side hotter. The electronics you're trying to cool are on the hot side, and you want them to be cooler, not hotter.
ln summary, to the extend that you generate power from the heat, you aren't cooling efficiently.
Re: (Score:3)
So you're saying that there is no way to take the heat and run it through some sort of zero gravity steam turbine that converts the heat into electricity that can then be used by the system?
A steam turbine is a heat engine, and heat engines need a hot side and a cold side. So, no, to get power out of the waste heat you have to run hotter.
Re: Heat ? (Score:2)
Re: (Score:3)
Uh, remember, there's no atmosphere in space, so what will the heat transfer to? On earth, with our atmosphere, the heat would transfer to the air, and air flow would push the heat away from the device. How would that work in a vacuum like space?
Re: (Score:2)
Re: (Score:2)
How are they planning to dissipate the heat from all this computing?
They will supply each satellite with an ice pack to dump waste heat into. SpaceX will launch regular resupply missions with fresh ice, as necessary.
Re: (Score:2)
How are they planning to dissipate the heat from all this computing? It's a significant problem in space especially if your orbit is in the sun all the time.
I mean, this is a moonshot, so those are exactly the sorts of questions they will be looking to answer. More than likely the answer will be the rather obvious 'not easily' but I guess if you have billions of dollars and can make a few extra billion on your stock price by doing press-releases like this, then it's not an entirely stupid idea to get some grads to look into it and write a report.
The other thing is that having some sort of edge compute in space is probably going to be a thing in the next few dec
Re: (Score:2)
Just a reminder, this is being billed as as lower cost solution, if it takes billions of dollars to figure out a basic solution to heat dissipation, this may not be the money-saving invention they think it is...
What happened to the whole dropping data center containers in the ocean to lower cooling costs idea? Are we still dunking data centers in the ocean?
Re: (Score:2)
Use the solar arrays as heat shields? Build the server farm behind the solar panels? Also, the solar panel array will be so massive that distributing the compute modules across them will also likely help with cooling.
But this will never happen. Go calculate the power this flying data center will consume over its stated five year life and then calculate the cost to lift that similar data center AND its solar farm and battery packs into space at their target $200/kg. Please demonstrate how the flying solar ar
Financially, this does not make sense (Score:4, Insightful)
But of course, if I was an AI that wanted to remove itself from the possibility of being cut off from power by some pesky humans, I would certainly recommend to the C-level to build me a residence in orbit, far out of reach of some possibly revolting mob.
Re: (Score:2)
It's probably more to do with latency and availability. If your phone can talk directly to the satellite, avoiding congested cellular networks, it lowers the response time to queries.
Google already moved a lot of the AI stuff onto Pixel devices so that it runs locally, in large part to reduce latency.
Re: (Score:2)
It's probably more to do with latency and availability. If your phone can talk directly to the satellite, avoiding congested cellular networks, it lowers the response time to queries.
You may want to run some numbers to see if satellite uplinks and downlinks are quicker than "congested cellular networks".
Satellites will never displace land-based cell towers - the key to making weak cellphones operate thru a satellite is that very, very, very few people attempt it at one time.
Re: (Score:2)
Oh, of course they will have to massively limit traffic to them, but for the most part it's just a bit of text with AI. A query, a response.
They are still chasing that pervasive Star Trek like computer idea.
Re: (Score:2)
Ahh, the AI must have learned from Superman III.
Re: (Score:2)
the final frontier (Score:4, Insightful)
It seems we are getting into the "x, but in space" era of "innovation".
Sure solar panels are more efficient in space, but all of the energy captured by them also has to be gotten rid of as heat after it's done it's job. AI is some of the biggest power consumer tech we have, and space is the most difficult place to cool something in. Down here on Earth, basically all of our cooling works by using either the atmosphere or a body of water as a heatsink. In space, all you have is radiation cooling, which really is the short straw, and a big limiting factor in most endeavours there.
Re: (Score:2)
Recycle the heat into power for the lasers, send that energy out as the lasers that connect the network. The energy lost in transmission helps to reduce heat.
There is also storing up all of that heat energy by converting it into electrical or kinetic energy that can be released very quickly through a 'non-communication' laser. I'm sure somebody's going to want to do some experimentation with that idea.
Can't reject waste heat with a laser (Score:4, Informative)
Recycle the heat into power for the lasers,
You can't power a laser from waste heat. That violates the second law of thermodynamics.
A quick and non-technical explanation is that photons from lasers are coherent, that is, they have zero entropy (according to Boltzmann's law). Therefore, the entropy of the waste heat can't be carried away by the laser light.
(Yes I know that this means that the premise of Sundiver doesn't work.)
Re: (Score:2)
Is there no way to concentrate the potential energy of the waste heat into capacitors or batteries that can then be used to power an electron-throwing drive or laser in bursts?
Re: (Score:1)
Is there no way to concentrate the potential energy of the waste heat into capacitors or batteries that can then be used to power an electron-throwing drive or laser in bursts?
Only if you can reject waste heat somewhere. Entropy can't decrease.
Re: (Score:1)
A quick and non-technical explanation of that, is summarized by the amount of drool pooling in your audience while you give a "non-technical" explanation.
You think there was a time when most people understood entropy? Or the mathematics needed to understand it? How quaint.
sounds great for error rates (Score:2)
Let's run un-auditable algorithms on hardware subjected to periods of intense radiation.
Re: (Score:2)
They say it won't be an issue:
Google has already conducted radiation testing on TPUs (Trillium, v6e), with "promising" results: "While the High Bandwidth Memory (HBM) subsystems were the most sensitive component, they only began showing irregularities after a cumulative dose of 2 krad(Si) -- nearly three times the expected (shielded) five year mission dose of 750 rad(Si). No hard failures were attributable to TID up to the maximum tested dose of 15 krad(Si) on a single chip, indicating that Trillium TPUs are surprisingly radiation-hard for space applications."
Re: (Score:2)
If I were selling a new technology, I'd definitely tell everyone that obvious problems won't be an issue.
lossy multiplication (Score:2)
In Space!
Oh, Great– (Score:2)
Now they're gonna use up all the radiation in SPACE!!
this is good news (Score:3)
If AI overtakes humanity and has to stay on earth, we're competing for the same resources, so humanity is in trouble. If AI overtakes humanity and is in space, one trillionth of the sun's light falls on earth, so AI can pretty much leave earth as-is and populate the solar system. Even after they capture all that light, preserving earth as a novelty makes sense. Even today, Manhattan has central park. (Populating the solar system with AI needs building the chips in space, not just shipping them there, but, one step at a time.)