Ask HN: Cheapest option to colo a GPU compute box?
Hi everyone,
We need to build a GPU compute box (thinking GeForce GTX Titan Black[$1000] as the GPU) for training Convolutional Neural Networks and building services on top of the trained model. What would be the most cost effective way to colo such a server?
Hurricane Electric told us a 2U rack space (100mbit and 750W) would cost us $700/month and I find that a little too expensive.(they are saying that we would rent a whole cabinet for the same price since their main cost is going to be our power usage and not bandwidth etc)
I am also thinking of building the box on my own and putting it in our office in SF (which has a static ip and 50mbit unlimited internet) but that'd probably violate the TOS - and even if it doesn't I am not sure if the internet connection would be stable enough to operate such a service from (though it doesn't necessarily need to be real-real-time -- we just the box to consume our job queue and a 1 minute delay would be acceptable there)
What are your thoughts on this? Do we really have to come to terms with having to pay $700+/mo for this?
(somehow arranging this for ~$300/mo would be amazing)
5 comments
[ 4.5 ms ] story [ 20.3 ms ] threadOther option I'd look at is spot gpu instances on aws - then it's hosted and you don't buy anything - and potentially you could shut it down part of the time although you're charged a full hour for each partial hour...
But the thing is I am not sure if amazon keeps a spot-instance reserve (i.e. what happens if all the spot instances get turned into on-demand ones over, say, a month and suddenly your service is shut down and you keep waiting indefinitely[you'd probably wait for around 24 hours and then try to get an on-demand one - which is 11x more expensive, costing around $600/mo for a shared gpu. and if you can't find an on-demand one, you will be offline for around a week until you figure out a way of provisioning a gpu instance from another service provider, which will probably cost you a lot more, but at that point you will have to do it not to disrupt the service for so long])