Hello, I’m looking to build my own DL rig. I don’t mind getting top hardware (with reasonable bang for the buck). One area that bothers me is GPUs. It seems like graphics cards are overpriced at the moment (e.g., 1080 Ti regularly goes for over $1k when in stock). Given this card is almost a year old and almost 2x original MSRP, does it still make sense to get? I’m thinking crypto mining craze may not go away soon, so what’s a good alternative in today’s environment to still get by and take fast.ai courses and compete in Kaggle competitions?
You can go the AWS / some other cloud route and I have done this for quite a while. But nothing beats having your own box you can ssh into whenever and from wherever. There are so many aspects here, let me just say that qualitatively it feels much different, having your own box feels much nicer (no clicking buttons for spin up etc) and you can leave things how they are and be back to them whenever (I have a tmux session open).
I think the best piece of advice I can give: if you are far enough into the course and have trained multiple models and various datasets and feel that this is something you will continue with, then getting your own box is likely a great idea. Though depending on your level of confidence with maintaining your own system it might take you a while to bring it up to speed.
Up until everything I have done to right now I could have used any compatible GPU with 8 GB of RAM so maybe that is something you might want to start with. Then when you need more RAM you could just use this one as a secondary GPU to run small experiments when the big card is working on something. But I am not there yet where that would be something I would need.
As for the cloud environment - I think paperspace is the current provider of choice but I might be wrong on this one?
I was in a similar conundrum recently and after a lot of research, ended up getting a machine with a 1070 only because I got a lucky deal.
Look up pre-configured desktops - they are currently cheaper than the sum of their parts due to GPU prices and you’ll great support for a few years
It’s basically everyone’s doubt!
If you are into deep learning (and maybe will deal with crypto mining in the spare time), go for a personal box with a GPU with minimum 8GB of RAM like @radek says, and 16+GB of system RAM!
Do you have any reference to best pre-configured desktops.
kanish its better to use aws services.as you may have to maintain the desktop a lot.especially you have to check overheating problem in india
Hi, I got mine from ant pc. But recently, Acer has been slashing prices of laptops with nvidia 1060 to well below desktop levels so you can look out for those if you’re ok with a laptop instead of a desktop.
I haven’t had an overheating problem (yet) but having my own machine is a lot more convenient, there is the cost of the disk if you don’t want to download data every time you start a machine and the K80 is slower than my 1070. The thing I really miss is the download speeds on cloud machines.
What we forget to often mention along with the configurations is that you need bazooka internet speeds to download the Datasets for DL.