Comment by fooker
11 hours ago
Maybe we are talking about different things here?
I don't doubt spacex can fail at this.
I also don't doubt we are fairly close to making this plausible.
> plenty of them who have spent careers building data centers
Famously, plenty of people who have spent careers building rockets would swear that reusable rockets would absolutely never work.
>I also don't doubt we are fairly close to making this plausible.
Maybe you should doubt that. There's literally no reason to think this is plausible besides some hype merchants' say-so.
> some hype merchants
Excluding Spacex:
Nvidia, Google, China, European Commission, Blue Origin
And this being HN, a YC funded company has put a single GPU rack in space and demonstrated training a reasonable sized model on it.
But yeah, it's all hype, sure.
On the off chance you're sincere and not just heavily over indexed into Elon stocks:
It's trivial to understand why this is all hype if you pay attention to physics, as another commenter suggested earlier.
https://en.wikipedia.org/wiki/Stefan%E2%80%93Boltzmann_law
Assume you're radiating away the heat for a single B200 (~1kW), and the max radiator temp is 100C, you find A = ~3m^2.
So that's 3 square meters per GPU. Now if you take into account that the largest planar structure deployed into space is ~3k m^2 (https://investors.lockheedmartin.com/news-releases/news-rele...), you're looking at 1000 GPUs.
That's a single aisle in a terrestrial data center.
Cost to deploy on earth vs satellite is left as an exercise to the reader.
4 replies →