Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Starcloud

blogs.nvidia.com

Post reply on HN
Print view View original

201–210 of 234 posts

Re: Starcloud

#201

Post by nine_k »

phillipcarter wrote: post #120
nine_k wrote: post #60

Earlier quoted context omitted.

I think the bigger thing about a space-based data center that it's not on anyone's land, and not easy to inspect or capture. Solar energy available around the clock allows it to be self-sufficient for a long time. I suppose there will be some demand for high-security, high-price setups like that.

> that it's not on anyone's land Oh you can bet that, if we assume this happens in 10 years, various countries will absolutely do a "land grab" up high. There is no escaping it.

Space is no one's land by a number of active international treaties, and also very large and empty, so enforcing boundaries is hard, except by actively killing spacecraft up high. There is no viable "space defense", comparable to the atmospheric air defense. Were it not so, spy satellites won't exist.

Re: Starcloud

#202

Post by nine_k »

notahacker wrote: post #125
nine_k wrote: post #94

Earlier quoted context omitted.

The access is the customer's concern, much like starlink.

The customer is going to be extremely concerned when it turns out physically locating datacentres in space doesn't actually render the data inaccessible or uncensorable...

To render your data inaccessible, use /dev/null. For practical purposes, some access is required.

Censoring data in a datacenter in space requires either administrative access, or physical access. The latter is complicated in space, The former depends on your trust to the operator, and your security posture.

Re: Starcloud

#203

Post by xnx »

ryanisnan wrote: post #189
xnx wrote: post #186

Earlier quoted context omitted.

The Apollo passive radiators can dissipate ~2500 Watts into space. With most systems shut down, only ~500 Watts was coming from the remaining systems and the astronauts bodies.

Cool, thank you. So I read this as fundamentally, the heat they dissipated far exceeded the heat they produced. Do you mind opining on what similar figures would be with modest passive radiators and a typical data centre rack heat output?

No idea what the passive radiators might look like (50x the size of Apollo?), but an Nvidia GB300 NVL72 uses 120,000 watts.

Re: Starcloud

#204

Post by notahacker »

nine_k wrote: post #202
notahacker wrote: post #125

Earlier quoted context omitted.

The customer is going to be extremely concerned when it turns out physically locating datacentres in space doesn't actually render the data inaccessible or uncensorable...

To render your data inaccessible, use /dev/null. For practical purposes, some access is required. Censoring data in a datacenter in space requires either administrative access, or physical access. The latter is complicated in space, The former depends on your trust to the operator, and your security posture.

Since the admins aren't in space, actors that want to use administrative privileges to interfere with your data have no less access to it than if the datacentre was located on the ground.

The difference between the US government censoring a datacentre in orbit and one in California is a matter of cost rather than practicality, and it's actually easier for other spacefaring powers to interfere with it in a deniable manner if it's that important to them than the datacentre in California

Re: Starcloud

#205

Post by nine_k »

notahacker wrote: post #204
nine_k wrote: post #202

Earlier quoted context omitted.

To render your data inaccessible, use /dev/null. For practical purposes, some access is required. Censoring data in a datacenter in space requires either administrative access, or physical access. The latter is complicated in space, The former depends on your trust to the operator, and your security posture.

Since the admins aren't in space, actors that want to use administrative privileges to interfere with your data have no less access to it than if the datacentre was located on the ground. The difference between the US government censoring a datacentre in orbit and one in California is a matter of cost rather than practicality, and it's actually easier for other spacefaring powers to interfere with it in a deniable ma…

This depend on you threat model. If your model is mostly legal threats form less-than-nation-state actors, being formally outside any terrestrial jurisdiction may help. If you try to protect yourself from a big threat that won't mind raiding (or bombing) your DC without a court order, quite possibly locating it in space is not the best idea.

Re: Starcloud

#206

Post by lawlessone »

kibwen wrote: post #181
btown wrote: post #167

Earlier quoted context omitted.

In all seriousness, if AI models can handle quantization, they can handle some flipped bits from time to time! There are probably some fascinating papers to be written around how to choose which layers in an LLM architecture could benefit more than others from redundant computation in a high-radiation environment.

Brilliant, to turn up the model temperature we just hinge open the shielding. I call dibs on the patent!

Ok, has anyone patented chips with radioactive source glued to them? For "true" randomness.

If it not i want dibs on it.

Re: Starcloud

#207

Post by DontBreakAlex »

lawlessone wrote: post #206
kibwen wrote: post #181

Earlier quoted context omitted.

Brilliant, to turn up the model temperature we just hinge open the shielding. I call dibs on the patent!

Ok, has anyone patented chips with radioactive source glued to them? For "true" randomness. If it not i want dibs on it.

https://en.wikipedia.org/wiki/Hardware_random_number_generat...

> and even the nuclear decay (due to practical considerations the latter, as well as the atmospheric noise, is not viable except for fairly restricted applications or online distribution services)

Re: Starcloud

#208

Post by caminante »

torginus wrote: post #93

Altman: has stake in nuclear power and AI companies Also Altman: Let's build gigawatts of nuclear for AI Musk: has stake in space and AI companies Also Musk: Let's build AI datacenters in space

/uj Apparently, this has already been discussed a few times on HN, last year. The bubble is still expanding!

Re: Starcloud

#209

Post by gs17 »

MrZander wrote: post #197
runlevel1 wrote: post #179

Their numbers strike me as very optimistic: *Table 1. Cost comparison of a single 40 MW cluster operated for 10 years in space vs on land.* | Cost Item | Terrestrial | Space |:------------------------------|:--------------------------------|:---------------- | Energy (10 years) | $140m @ $0.04 per kWh | $2m cost of solar array | Launch | None | $5m (single launch of compute module, solar & radiators) | Cooling (chill…

> $5m (single launch of compute module, solar & radiators) This seems absurdly low to me.

It is, unless you take Musk's hype about Starship as fact. With rockets that are actually potentially available the best price is $1500/kg to LEO, so either they're presuming the whole setup weighs in at 3-4 tons (which is less than the shielding alone) or that they can get it launched for a few orders of magnitude less than what's on the market now (and they do say they assume $30/kg).

Re: Starcloud

#210

Post by asimpleusecase »

I could see gov imaging satellites with a direct encrypted laser communications to GPU’s in orbit being attractive. Images processing, movement pattern analysis, multi spectral, and as they mentioned radar.
Post reply on HN