undefined | Better HN

0 pointsNitpickLawyer3mo ago0 comments

> They put the hardware inside,

Which is exactly how you'd do a hypothetical dc in space. Come on, you're arguing for the sake of arguing. CotS works. This is not an issue.

> That kind of shielding is going to blow your mass budget

SpX is already leading in upmass by a large margin. Starship improves mass to orbit. Again, this is a "solved" issue.

There are other problems in building space DCs. Rad hardening is not one of them. AI training is so fault tolerant already that this was never an issue.

0 comments

adastra223mo ago

None of the discussed designs include radiation shielding like that. Nobody is considering doing it that way, because the math really really doesn’t work out (instead of unshielded, where it just doesn’t work out).

A cosmic ray striking a chip doesn’t cause a bitflip - it blows out the whole compute unit and permanently disables it. It is more like a hand grenade going off.

notrealyme1233mo ago

> AI training is so fault tolerant already that this was never an issue.

Such nonsense.

NitpickLawyerOP3mo ago

Between fp nondeterminism, fp arithmetic, async gradient updates, cuda nondeterminism, random network issues, random nodes failing and so on, bitflip is the last of your concerns. SGD is very robust on noise. That's why it works with such noisy data, pipelines, compute and so on. Come on! This thread is having people find the most weird hills to die on, while being completely off base.

j / k navigate · click thread line to collapse