Hacker News Reader

Comment by a11r | Hacker News Reader

That whitepaper puts the cost of a 40 MW cluster at $167m on land and $8.2m in space, inclusive of launch costs.

They seem to not be counting the $1B of GPUs on either side, making the difference pretty small. They also think launching 200 racks of equipment into space will cost $5M.

that sounds ridiculous. How could it be so much cheaper

Because launch costs and density are completely imaginary. The whole thing is predicated on launch vehicles they hope someone builds eventually. They also appear to have profoundly miscalculated their radiation shielding costs and amounts, but I'm not really motivated to calculate that correctly.

At planet we didn't have any radiation shielding. Why would spacex need it?

On the planet we have tremendous radiation shielding. The van allen belt and the atmosphere, to start with.

I suppose “at planet” refers to the company Planet

https://www.planet.com/products/planet-insights-platform/

In LEO you would still have the Van Allen belts.

Because running 40MW of high-speed inference on GPUs is a different sort of mission then bolting a camera to a Jetson SBC and dropping a cubesat into orbit?

Why is it different. I don't understand at all. I thought this was all shown to not be needed since even before we did it at planetlabs

They're assuming $30/kg launch costs.

Falcon heavy has never flown SSO, but it's probably over $1500/kg, which is >50x more expensive. Even the most optimistic estimates for a future starship launch cost is at least $100/kg, so it makes no sense to use $30/kg.

The other line items are probably similarly unrealistic since they include their own launch costs as well.

They also spec 300 racks of 500kg each, which is <40% of the weight of the comparable land rack. I doubt the full weight reduction since the racks still need to survive launch force/vibration in addition to fluid/pressure enclosures that land racks don't require. I think it's likely that just about every part of this report has a similarly cooked numbers, and we can just throw the whole thing out.

Because they have factored in a genie’s lamp, which they will use to wish costs down significantly

how much area do you need to radiate away 40 MW, assuming you want to keep your GPUs at a toasty 90 celcius. Every space datacenter fan/promoter enthuses about the free solar power, but forget to mention how the heat will be dissipated.

Is there not enough space in space?

Space, as in the near absence of material that makes it an almost perfect insulator. Add to it constant compute producing heat and you just cooked your GPUs with no way to dissipate.

That must be why everyone on the ISS has died of heat stroke.

You surely walk around with fans all around your body similarly to GPUs, even though you are in the same environment as them?

If not, then you may realize how stupid your comment is.

I don't know why people are so fixated on heating as some kind of gotcha. If we already assumed magical 16km^2 solar array, then we can also assume matching size radiator, with heatpumps. If ISS can do it, then there is no technical issue with doing it on bigger scale. It "only" doesn't make any sense from financial viewpoint, terrestrial datacenters beat it in every way.

NVidia's recent support for warm (45 °C) water cooling for their new racks probably had more impact on price of compute than than space servers will ever have.

Reply on news.ycombinator.com