On the other hand: the heat has to go somewhere. So… where? Datacenters already create a warm microclimate in their vicinity, is that getting even worse?
It’s kind of like how brine from desalination is not a global problem for the oceans at all — all that matters is diluting it enough that it doesn’t poison the local ecosystem.
Nvidia has so much money and they can’t afford to pay a human for a day of their time to write a blog post?
We are all fucked.
And it’s sad because Jensen seems like one of the rare good CEOs when I listen to him speak.
But even Dario says he doesn’t let Claude actually write his blog.
Have we been listening to the same person speak for the last few years? Jensen rarely even sounds sane anymore.
The same as technical docs for any codebase, humans will not read them anymore, only AIs which then translate it to human on-demand, it's already happening, I've worked recently with many new frameworks/codebases without even opening the doc (not even the Github page) and solely asking the agent to gather info for me about it.
PS: The reason I feel it will be this way is that it will allow to legitimatize mass data collection indirectly, instead of doing telemetry on page and software level, we will just send all the content automatically to some inference providers (probably provided for free by Google, MS and so-on)
Summer is still an issue, but fun solutions are possible. With the right geology, I think it’s possible to heat an underground volume in the summer and recapture (some of) that heat in the winter. In many, many climates, annual heating costs are far higher than cooling costs, at least if people aren’t stupid with skylights. [0]
[0] As a back-of-the-envelope heuristic, heating or cooling load due to conduction and air exchange is proportional to the difference between indoor and outdoor temperature. Outdoor temperatures of -10F to 30F are not unusual in the winter and are 40-80F away from an indoor temp of 70F. But outdoor temperatures in these climates rarely exceed 95F and are mostly lower in the summer, so that’s 15-25F of cooling. And heat pumps are more efficient at smaller temperature differences.
Radiative heating is an entirely different story.
I do think ground based centers will have better economics when they can be built though, and this addresses noise and water complaints which are the big 2 regional complaints.
It seems like lots of bottlenecks are getting solved quickly, except for maybe memory.
To be precise, heat rejection via radiative cooling scales with the fourth power of the temperature (in K) the radiator operates at, all else constant.
What's a favorable climate, apart from, obviously, Greenland? The piece is a little light on details on the correlation between outside temperatures and efficiency & cost. It'd be nice to see even a broad-strokes discussion of that.
https://www.kit.edu/kit/english/pi_2024_038_kit-supercompute...
I assume Germany is the same, many years ago really is different to today.
https://www.nasa.gov/centers-and-facilities/ames/doing-more-...
https://www.nas.nasa.gov/assets/nas/pdf/ModularSupercomputin...
Or are they for some unfathomable reason using evaporative cooling in data centers?
The loss of material must be included
If water is evaporated or spent out of the system.. it is not more efficient
I get that they're using liquid coolant at higher than usual temperatures, but why couldn't they do that before? Most of the comparison in the article is for air cooled datacenters but what about other liquid cooled ones?
Surely in all the previous datacenters that have been designed there has been someone doing the math and determining what temperature things need to run at, how much energy it will use, how much heat it all will produce, etc.
edit: just saw this:
>Previous liquid-cooled servers were hybrid: GPUs and CPUs got cold plates, but the rest of the system stayed air-cooled, with finned heat sinks designed to shed heat into moving air. In a fully liquid-cooled server, the cooling for these components needed to be completely redesigned to use liquid.
It's not clear to me what changes are happening here. The siblings to your post seem to be indicating an overall improvement.
If you manage to use the waste heat to avoid generating heat somewhere else (that the article calls heat recovery) then there’s a further reduction in total heat output.
> Nvidia has so much money and they can’t afford to pay a human for a day of their time to write a blog post?
The shareholders desperately need that money.
With that said, by the standards of industrial sites data centres are quiet, low traffic and smell free. An industrial area that can’t build a data centre certainly can’t build a steelworks or oil refinery or leather tannery.
No one bats an eye when an air conditioner runs.
In the US
Power plants are all over, even in populated areas. They’re not so bad either (except perhaps coal).
There is no fundamental reason that datacenters need to be especially unpleasant to their neighbors.
I don't have any datacenters near me but I can hear some heavy hums from the washing machine 3 floors up when it put my head on my pillow, for some reason it just propagates through the building physically. When I walk around I don't hear it. Datacenter noise can be the same.
IMO they should be put away from habitation, there's no reason for them to be near there anyway
I agree, but that's a hard problem (in the US anyway). Unless you're plopping data centers in the middle of national parks, or in the middle of the desert where water is going to be a problem, you are nearly always going to be within some small mile radius of civilization. Plus the cost of trenching new fiber out in the middle of nowhere.
The same reasons humans want to concentrate in a particular area (access to jobs, infrastructure) are the same things that data centers need.
Once water-less cooling tech like this improves then yeah, just plopping them in the middle of the unpopulated desert becomes viable (assuming you can get the fiber out there and latency is tolerable), so long as they generate their own power.
Imo we should just solve the problems with data centers being near cities. Manage/regulate the noise and any waste (heat included, it shouldn't drastically impact the neighbors) and make them pay for any utility capacity/reliability upgrades needed. If this article is right and water usage can be nearly eliminated then it seems like the rest should be solvable? Especially if we can take the extra heat and use it for local power or heating needs.
If you cycle between 45 C and 55 C water temperature (as mentioned by the press release), you are only getting a 10 C delta. That isn't even enough for district heating, probably not even with heat pumps.
Now if you have something like a steel foundry, that have much hotter cooling water, you can absolutely use the heat for district heating, but even then it usually isn't enough for cost effective electricity generation. Even when it is waste heat, as the equipment to handle it still costs money and requires maintenance.
Right. Vibration propagates through solid (and liquid) materials.
But this can all be measured and controlled, and there's nothing special about datacenters. A building that is hundreds of feet away will couple to your pillow much less strongly than a washing machine in your building. And the washing machine often has a wildly unbalanced load and minimal decoupling between itself and the floor, whereas a big fan in a datacenter or other industrial building ought to be balanced and also ought to be installed on decoupling mounts.
If datacenter operators (cough xAI) are being lazy about properly selecting, installing and maintaining equipment, then you can have a problem. Otherwise you have a much smaller problem.
People said this about high voltage electric lines and wind turbines. Blind tests proved they were imagining things.
Sure there is, being a good neighbor costs more than being a bad neighbor
So what, winters would be no more? Snow will disappear, no more ice-men and christmas trees, and subzero conditions in general, too?
You do eat, don't you?
I think we are going to need heating.
I don't live next to one but I'd take constant humming over the constant stop/go traffic noise, honking, squeaky brakes, slamming doors and revving engines I now have on my western side of the apartment, thanks to the unemployment office the city opened on my street not too long ago.
So how come constant humming is somehow an illegal nuisance, but we've been expected to put up with the much more annoying urban traffic noise for decades just fine?
My parents apartment have constant humming anyway thanks to the HVAC system on the roof of the nearby supermarket and white/brown noise is far more tolerable and easy to tune out than traffic noises.
https://xcancel.com/BrianEntin/status/2067930868191035474?s=...
For one, there tends to be little traffic at night when most people want quiet in order to sleep. Driving is also something (nearly) everyone does and benefits directly from, so negative externalities are easier to accept. It is much harder to accept a new source of noise near your home you haven't asked for and don't directly benefit from.
(Seasonal heat storage is also a thing, Espoo's neighbours have tens of GWh of storage, with a new 90 GWh cavern in the works. Not sure if the systems are interlinked.)
Then 45 or below is sent back on the return.
DC inlet is 45°C, outlet is 55°C assuming a 10°C ΔT. By the time that's travelled 500m–1km through pipework you've lost a few degrees, so you're arriving at the HIU at maybe 50–52°C. The home radiator circuit then takes that down by around say 12°C, returning ~38°C. Factor in pipe losses on the return leg and you're back at the data centre with maybe 35°C inlet rather than 45°C — meaning the DC output is now only 45°C rather than 55°C, and the whole system gradually degrades each cycle. You could address this by mixing some hot output back into the return to keep the DC inlet stable at 45°C, but eh.
I mean people make clusters out of raspberry pis and minipcs.
Maybe look at what is inside a datacenters, the amount of power required is very large, and the hardware to run the inference + network isn’t small. Then try to see how much sending that to space cost
Yes
> And that can't ever change?
It can, but but not for free. As the comment earlier in the thread was referring to, more computing power needs more cooling, and cooling in space is hard.
> A reasonable person doesn't understand technology usually.
What? Do you mean a layperson? Why does that matter when discussing the feasibility of space-based AI datacenters?
> That is increasingly an understanding left to the wizard class.
No, you can get there with a bachelor's degree in a relevant subject. Or just reading informative news sources.
> I mean people make clusters out of raspberry pis and minipcs.
So? What does that have to do with anything?
And yes, a space-based computing node would not need quite as much of some of these things but they'll still need them in some way. It's not like you can just plug in a power and ethernet cable into them.
I doubt this will scale to a level that is actually useful. It's a nice experiment, just like Microsoft when they threw a datacenter container into the ocean. But not practical in the current conditions: https://news.microsoft.com/source/features/sustainability/pr...
Yes they say it is amazing and sustainable there in that blog post, yet somehow they've never bothered to do it again.
https://i.redd.it/zh7qvyfqgvx21.jpg
So to me they have solved the issue of having a space based compute array network interfacing with the earth. They have solved the issue of launching and deploying this array. And their given launches seem to have a substantial payload of compute going up at once just in sheer volume. And right now the only real difference is that the nodes they are launching are just pretty weakly specced. Everything else is in place and turnkey.
They are sending a few racks of stuff up but the problem is not that it's underspecced. It's that most of it is just needed for survival and communication in space.
You're talking about an environment that's full of radiation and goes from -200C to +200C every 90 minutes. That needs to be orbit managed and cooled without any airflow. Just sticking a few servers in a barrel isn't going to do the job.
I do really think that if large numbers of jobs are indeed going to be displaced by AI, movements will pop up of people attacking datacenters (and honestly I wouldn't blame them even though it won't really accomplish anything). Having them in space keeps them out of reach of anyone but state actors.
If we pick an extremely fast orbit, then relativity means the hardware will age out (slightly) slower, so I'm sure that'll help with the maintenance issue.
It's the wrong way around though. Ideally we want to speed up our current compute ability not slow it down; if it experiences more time than we do then it can do more. Relative-MHz means my slower hardware becomes tangibly fast again.
General Relativity says mass warps space time, so we need to get these datacentres out of the Earth's gravity well. And the Sun's, and the Milky Way's; out into the deepest void of intergalactic space. The good news is that a maintenance callout is still quicker than some of the earth based DC's I've had gear in, but the bad news is that it doesn't get us much of anything at all.
Special Relativity lets us abuse time with speed (something I discovered as a teenager). Going faster than Earth means we experience less time, so we just need to try and slow down comparative to our home base. The earth is orbiting the Sun at ~30km/s, the solar system is orbiting the centre of the Milky Way at ~230km/s and our local group of galaxies is moving relative to the Cosmic Microwave Background at ~600km/s. We can easily get our DataSpaceCentre up to 1,000km/s or more, so we just need to point it relative to all that movement we mentioned above making stationary relative to the universe. It's completely doable, but (as well as far more variable response times to callouts) only gets us an extra second of compute over a human lifetime.
Fundamentally, we're attacking this problem in the wrong direction. Earth's gravity is comparatively minor, and our piddly ~600km/s relative movement is a tiny fraction of the speed of light. We should be filling The Earth with compute, and then decamping humanity into space and travelling at relativistic speeds. Or put the compute in space and move the Earth into the event horizon of a black hole. You can't do the inverse of Interstellar keeping Earth where it is, the maths isn't in our favour. If everyone lived on (a less moist) Miller's Planet, we'd get 7 years of compute every hour. It puts Moore's Law to shame; the relative MHz are obscene.
There's the obvious problem of communications. I'm led to believe there's issues with radio and light, so this probably isn't a job for fibre. Veritasium seemed to imply a battery, switch, lightbulb and a wire stretching around the globe would light instantaneously, so I'm sure we can come up with a new copper Ethernet standard for low latency over solar distances.
Invest early, we're going straight past the moon!
https://spacenews.com/china-backs-orbital-data-center-startu...
https://www.reuters.com/science/china-vows-develop-space-tou...
Plans for space data centers should be seen with skepticism. However when they are backed by different parties who have stakes in the game, that's more credible. More than HN crowd for sure.
1. Space is terrible for heat regulation. It's a perfect insulator for everything except radiative cooling, which is the least efficient. Hot things stay hot.
2. Space is full of radiation. Everything has to be radiation hardened, which makes it heavier, more expensive, and, yes, more difficult to cool.
3. Space is far away. Well, farther than a data center on Earth can be. I know China hasn't solved the speed of light.
We put up with it with satellites because it still has some advantages over trying to run cables literally everywhere, but we do, in fact, still use cables laid on the bottoms of the oceans.
So, is physics wrong, or is a country known for making dumb decisions some times making a dumb decision?
For e.g you might think of the outskirts of London as fairly moderate, but this week it's been hot enough that supplemental cooling would likely have been needed at points. For a data centre here you'd typically design the cooling system to cope with outdoor temps in excess of 40°C, which is not a conservative number anymore.
Also, while Nvidia might be happy with you supplying water at 45°C I suspect you will get better longevity of the hardware at lower temps like say 35°C. GPUs are expensive, and extending longevity may well be 'worth' a bit more water or energy to you. In practice you are also likely to have air cooled systems that sit 'beside' the AI compute like storage severs, any extra CPU compute and network switches. So you are likely to need a separate room and cooling system for that. Great progress though.
currently.
and also more energy efficient, because evaporating water away takes a lot of energy with it. you have to raise radiators to a higher temperature to keep up with that, or have much more surface area.
nialse•8h ago
eqvinox•8h ago
quickthrowman•7h ago
There are some systems that pipe refrigerant around the building, but they’re relatively uncommon (VRF or variable refrigerant flow if you want more details).
Glycol and water is cheaper than refrigerant so there’s usually a chilled water loop that passes thru a heat exchanger that interfaces with a chiller (vapor compression refrigeration) to reject the heat from the chilled water loop.
This eliminates the need for evaporative cooling towers.