Critique my DL box

Part 1 (2019) Advanced (Part 1 v3)
1

I am thinking of building the following box to use for fast.ai. Would really appreciate some comments:

https://ca.pcpartpicker.com/user/gilaniali/saved/#view=nKkvf7

  1. Is this ok for fast.ai
  2. Will this allow me to upgrade to a second GPU in a year or so?
  3. Should I install Windows as the base and use the Ubuntu OS app to run fast.ai or install Ubuntu as base and run Windows for games in a VM?

Thanks!
Ali

2

You’ll probably want to get a bigger power supply. 650W for 2 GPUs (plus your regular CPU) is a little tight.

With a 1 TB SSD, you’ve got plenty of space to dual boot between Windows and Ubuntu (i.e. install each on its own partition).

The motherboard does not appear to have WiFi, in case that’s important for you.

2 Likes
3

Wow didn’t even realize it didn’t have wifi - thanks! will look into a card or a better motherboard.

Dual boot is an option but if I am going to play games occasional VM might be a better choice. But lets see if performance degrades then Dual boot it is.

4

I think a lot of these cheaper mobos don’t have WiFi. You can always get an external WiFi dongle and plug it into USB. (Although really you’d want to plug this box into Ethernet for downloading those biiiiig datasets. :wink: )

5

Should look into connecting 2 GPU for faster data transfer: https://www.nvidia.com/en-us/data-center/nvlink/.

Personally I have never found the CPU to be a bottleneck, so I would always just throw in an old compatible one.

Have you already been through the course? I don’t think buying a machine specifically for DL makes sense until you have done a bit of deep learning. Your hardware bottlenecks would also depend a lot on what tasks you are doing.

For example I have done some training with GANs, and two 2080TIs instead of two 2070S would probably have have given me ~3x speed up. I was heavily GPU memory constrained, so I had to use a small batch size which gave me low utilization, with Gradient checkpointing for speed/memory tradeoff.

6

That’s a good point. I bought a new DL box 3 years ago for about $2500. It has been convenient but I also didn’t use it as much as I thought / planned I would. Now the GPU is out-of-date.

If I had used the cloud for all my projects, it probably would have been cheaper, especially since this includes GPU updates “for free”.

Then again, if you also plan to play games on this computer it might be worth the $$$.

7

I haven’t done the fast.ai course / book yet but investing in a machine will make me more likely to stick with it and learn more about deploying.

8

It definitely allows me to play games and work on side projects / use as a spare work pc etc. The 2070 GTX should be good for another 3 years i think before i would need to upgrade.