Comment by ianhowson
9 months ago
I had a good laugh.
- You can't build 40MW of solar panels for $2M, even with theoretical maximum efficiency. You can't even build the cabling and regulators at that price.
- You need battery storage -- not as your backup -- but as primary source. It is going to cost more than $2M. Batteries are heavy. They are going to cost a lot to launch. This is not even solved on the ground yet.
- You need a heat transport medium to move heat into your massive radiator. Either you use water or you use air or you use heatpipes (metal). You have to pay for the cost and weight and launch expense. This is probably half the weight of the rack and I haven't bothered to do the math about how you transport heat into a 500 foot solar sail.
- Let's not even talk about how you need to colocate multiple other racks for compute and storage. There aren't any 1TBps orbital link technologies.
- Rad shielding? It doesn't work, but I'll let this slide; it seems like the least problematic part of the proposal.
- 15 year lifetime? GPUs are obsolete after 12 months.
I don't want to be the guy who shoots stuff down just for fun, but this doesn't even pass the sniff test. Maybe you can get 10x cheaper power and cooling in space. Still doesn't work.
Also: repairs. Every time I read someone’s story about large-scale ML training, a bunch of it is about identifying failing or flaky equipment and fixing it. That’s not so easy in space.
Nonsense, it's right there in the acronym: Space Reliability Engineer (or I guess one could also just leave "Site" as is, since space is for sure a site). That PagerDuty rotation is gonna be hell
At this point as long as you say AI 3 times someone will give you money.
I guess you need connections too, and maybe a previous exit.
This idea in particular doesn't make any sense... Currently. Maybe in a decade or so with better technology.
Although the prospect of polluting the stars itself with a bunch of computers generating AI slips... We paved Paradise to put up a parking lot
It makes sense if you want to put a nearly unkillable AI in orbit to control life on earth :D
Where do I buy some of these 12 month old obsolete GPUs?
eBay works for me
What are you searching for?
Either this is performance art, or interest rates are too low.
Or it's a cover for putting something else up in orbit secretly.
Good point, Rad shielding … how even trust your calculations when everything is grilled by charged particles.
GPUs are not obsolete after 12 months. Look at how Nvidia is stagnating for their 50 series lineup.
The biggest problem is software. The CUDA stack is not maintained forever and certainly less than 15 years.
If you have continuous sunlight, can't you get away with no battery?
Not arguing with your overall point - this company looks ridiculous.
For continuous you need to either go for a polar orbit or go very far in space. Most launch centers & providers are not well situation for polar orbits because its not a common use case, so you need to sacrifice launch mass. The same goes for far away orbits - you need to sacrifice launch mass to go further. Also if you are far then you get latency issues.
So it skews the economics pretty harshly. I think OP is right - you need good batteries somehow.
I think the proposal suggested an orbit where the solar panels are always in sun and always properly aligned and always clean due to space gophers.
But more seriously, GPU loads are super spiky. Ground-based power grids and generators and batteries have trouble keeping up with them. You can go from 1MW idle to 50MW full power in 10ms. Unbuffered solar cells are right out.
1 reply →
> You need battery storage
How so? Is it not possible to position the satellite in an orbit that keeps it in perpetual sunlight?
I thought rad shielding works?
Just that it tends to involve heavy AF materials like water
Most comments on this page are about the problem with heat. You're saying the problem is battery storage.
... couldn't you just merge both problems into a solution - your radiators ARE you power source
Temperature isn't a power source; heat flowing across a temperature gradient can be. But that brings us back to the first problem - how to make it flow.