vrighter comments on An in-space construction firm says it can help build massive data centers in orbit

[–] vrighter@discuss.tchncs.de 2 points 1 year ago (5 children)

for starters, at the loads they're running at, they have literally hundreds of gpu failures a day. How do you propose doing that in space?

permalink

fedilink

source

parent

hideshow 5 child comments

[–] FaceDeer@fedia.io -2 points 1 year ago (4 children)

Include spares.

I hope they're reading this thread and taking notes, they probably didn't think of that.

permalink

fedilink

source

parent

hideshow 4 child comments

[–] vrighter@discuss.tchncs.de 1 point 1 year ago (3 children)

and the infrastructure and robotics to replace them, of course.

Assuming 200 nvidia H100 failures a day (conservativo, reality is worse) that's an extra ~340kg of weight you'd need to launch per day. Which is an extra 120 tons yearly.

permalink

fedilink

source

parent

hideshow 3 child comments

[–] FaceDeer@fedia.io 1 point 1 year ago (2 children)

So, one Starship launch per year. Doesn't sound like a problem.

permalink

fedilink

source

parent

hideshow 2 child comments

[–] vrighter@discuss.tchncs.de 1 point 1 year ago (1 child)

at least two, you can't stuff a rocket full of just gpus, you need something to actually dock and deliver the payload in space. So you need to launch at least 2 rockets (in a non-reusable configuration, so you need to pay for the whole rocket and the launch) to ship a bunch of gpus that are, at best, only 10% as fast as usual.

permalink

fedilink

source

parent

hideshow 1 child comment

[–] FaceDeer@fedia.io 1 point 1 year ago

Okay, two then. It's still cheap.

in a non-reusable configuration

Why do you say that? 120 tons is well within Block 4's projected capacity in reusable configuration. 240 tons is almost within it, even.

permalink

fedilink

source

parent