Last big crypto crash I bought a used Titan XP for like 350$ (still running strong and kicks ass). Got me through this whole GPU debacle. Always buy low!
I'm just finishing up Andrew Ng's course, and would like to pick up a GPU for building models (probably focusing on U-net image models), do you have any references on what I should expect from trying to buy a used one like you did?
I know it's a pretty random question, but i honestly don't know anything about GPU's other than that my life would be a lot easier if I had one (I run an old Macbook Pro, and I have an old Asus laptop that runs a debian distro, and finally my very old gaming pc is a alienware alpha).
I am no expert, but my off the cuff thought is that buying a gpy and trying to run it as an external for any of those machines will be pretty pointless.
First it probably won't work; certainly for the mac you will face driver issues that were beyond me the last time I tried (about three years ago). I suspect things are worse now... Nvidia have not been helpful. Probably you could get a radeon working as an egpu, but for ML you really need Nvida and Cuda (unless you are seriously into homebrewing code).
But the other issue is that you need to feed the beast - that is you need to pump data in and out of the GPU. For that you need a fairly fast CPU and fast ssd, and fast interconnects to the GPU. You also need power, a good power supply that can keep everything running -- including the fans... which need to be in a good box so that there is good airflow.. or quite quickly you will have no GPU!
Maybe build a box like this with a cheap nvidia GPU, get it working so you know that you have a viable rig, benchmark it to see if you are bottlenecked on the GPU (long time since I did this, so I'm not 100% how - but basically if the GPU is 90% for ages and everything else is sitting looking at it then that's a good sign). Then buy a top notch second hand (about £500 right now on ebay) and resell the cheap one. If you hit a hitch inbetween you know you have a problem to solve and haven't spent all your money. The extra points are for getting multiple GPU's running... again I did that in a corporate with supermicro boxes about 8 years ago and it drove me up the wall at the time... I suspect now it's possible to have 2 GPU's in a consumer box/motherboard but I am not sure what the state of play is.
Are you concerned about damage from heat during previous use? You're going to be running it for hours or days anyway if you train on large complex datasets, so a GPU that's been used for a few months might not have lost you too much lifetime.
