Cooling isn't anymore difficult than power generation. For example, on the ISS solar panels generate up to 75 W/m², while the EATCS radiators can dissipate about 150 W/m².

Solar panels have improved more than cooling technology since ISS was deployed, but the two are still on the same order of magnitude.

▲

Nevermark 20 hours ago | parent [-]

So just 13.3 million sq. meters of solar panels, and 6.67 million sq. meters of cooling panels for 1 GW.

Or a 3.651 km squared and 2.581 km squared butterfly sattelite.

I don't think your cooling area measures account for the complications introduced by scale.

Heat dissipation isn't going to efficiently work its way across surfaces at that scale passively. Dissipation will scale very sub-linearly, so we need much more area, and there will need to be active fluid exchangers operating at speed spanning kilometers of real estate, to get dissipation/area anywhere back near linear/area again.

Liquid cooling and pumps, unlike solar, are meaningfully talked about in terms of volume. The cascade of volume, mass, complexity and increased power up-scaling flows back to infernal launch volume logistics. Many more ships and launches.

Cooling is going to be orders of magnitude more trouble than power.

How are these ideas getting any respect?

I could see this at lunar poles. Solar panels in permanent sunlight, with compute in direct surface contact or cover, in permanent deep cold shadow. Cooling becomes an afterthought. Passive liquid filled cooling mats, with surface magnifying fins, embedded in icy regolith, angled for passive heat-gradient fluid cycling. Or drill two adjacent holes, for a simple deep cooling loop. Very little support structure. No orbital mechanics or right-of-way maneuvers to negotiate. Scales up with local proximity. A single expansion/upgrade/repair trip can service an entire growing operation at one time, in a comfortable stable g-field.

▲

pfdietz 10 hours ago | parent | next [-]

Solar panels can in principle be made very thin, since there are semiconductors (like CdTe) where the absorption length of a photon is < 1 micron. Shielding against solar wind particles doesn't need much thickness (also < 1 micron).

So maybe if we had such PV, we could make huge gossamer-thin arrays that don't have much mass, then use the power from these arrays to pump waste heat up to higher temperature so the radiators could be smaller.

The enabling technology here would be those very low mass PV arrays. These would also be very useful for solar-electric spacecraft, driving ion or plasma engines.

▲

krisoft 13 hours ago | parent | prev | next [-]

> active fluid exchangers operating at speed spanning kilometers of real estate, to get dissipation/area anywhere back near linear/area again

Could the compute be distributed instead? Instead of gathering all the power into a central location to power the GPUs there, stick the GPUs on the back of the solar panels as modules? That way even if you need active fluid exchanger it doesn’t have to span kilometers just meters.

I guess that would increase the cost of networking between the modules. Not sure if that would be prohibitive or not.

	▲	Nevermark 6 hours ago \| parent \| next [-]
		> Could the compute be distributed instead? For electrons that would dramatically increase latency, and lower bandwidth, slowing down compute. Maybe dense optical connects could work?
	▲	pfdietz 10 hours ago \| parent \| prev [-]
		It's easier to shield the GPUs if they're all grouped up.

▲

jaywee 12 hours ago | parent | prev | next [-]

Well, divide et impera. Fairly straightforward for AI inference (not training): The existing Starlink constellation:

3491 V1 sats × 22.68 m² = 79176 m²

5856 V2-mini sats × 104.96 m² = 614 646 m²

Total: 0.7 km² of PERC Mono cells with 23% efficiency.

At around 313W/m² we get 217MW. But half the orbit it's in shade, so only ~100MW.

The planned Starship-launched V2 constellation (40k V3 sats, 256.94 m²) comes out at 10 km², ~1.5GW.

So it's not like these ideas are "out there".

	▲	Nevermark 6 hours ago \| parent [-]
		Take those 40,000 satellites, and combine their solar panels, and combine the cooling panels, and centralize all the compute. Distances are not our friend in orbit. Efficiency hyperscales down for many things, as distances and area scale up. Things that need to hyperscale when you scale distance and area: • Structural strength. • Power and means to maneuver, especially for any rotation. • Risk variance, with components housed together, instead of independently. • Active heat distribution. Distance is COMPOUNDING insulation. Long shallow heat gradients move heat very slowly. What good does scaling up radiative surface do, if you don't hyperscale heat redistribution? And you can't hyperscale heat distribution in 2D. It requires 3D mass and volume. You can't just concatenate satellites and get bigger satellites with comparable characteristics. Alternatives, such as distributing compute across the radiative surface, suffer relative to regular data centers, from intra-compute latency and bandwidth. We have a huge near infinite capacity cold sink in orbit. With structural support and position and orientation stabilization for free. Let's use that.

▲

withinboredom 13 hours ago | parent | prev [-]

Lets not forget that you have to launch that liquid up as well. Liquids are heavy, compared to their volume. Not to mention your entire 'datacenter' goes poof if one of these loops gets frozen, explodes from catching some sunlight, or whatever. This is pretty normal stuff, but not at this scale that would be required.