Building home GPU server

Hello course mates,
I’m building a home desktop server for deep learning, to avoid costs in the cloud in medium/long term.

I have chosen to buy 3x GTX 1070, which seem nice in terms of cost/performance, but I’m not sure about a detail: does having a motherboard which allows the 3 graphics card to run in PCIe 3.0 x16 make a real difference? The reason I ask this is that I can get a very decent asus motherboard for having the 3 cards, but it doesn’t allow them to run at x16 at all. Probably they will run at x4 or x8 (not sure)… Can I expect a huge decrease in performance?

Thanks in advance for any feedback!

And thanks Jeremy for the hard work in the course! :slight_smile:


I’d suggest participating in the home server thread on the #part1 forum, since this has been well covered there. BTW it’s probably better to buy 1-2 1080ti’s instead of 3 1070’s.

1 Like

I’ve found my answer there almost instantly…sorry! And thanks for the feedback!

Hey you never need to say “sorry” for asking a question :slight_smile: We’ll all learn our way around these forums as the course progresses…


@Jifu I was lurking in the home server thread myself, and I looked at a few blog posts (@brendan’s, Tim Dettmers’). Folks have put up their DL system builds on PCPartPicker. These resources can keep one busy for a week!

Two things I took note of (for when and if I try to make a home box) was (a) make sure the motherboard has enough widely spaced PCI slots (16lanes) to support adding more GPU’s later and (b) amortize GPU cost by waiting for next version while you picking enough skills to make effective use.

1 Like

If you are using Windows 10
Ubuntu only works if you have Windows 10 Pro

Can someone with these…

I can’t find that mentioned online in a quick google search - do you have a source for that?

I tried repeatedly to install it with my Windows 10 Home but it did not work until I installed the Windows 10 Pro.

Maybe it was something related to the incorrect OS Build like it says on the
MSDN - Install the Linux subsystem

Maybe I paid 99.00 + Tax for nothing (Darn it!!)

Hey it’s not for nothing - now you’re a pro!


@Gerardo I was able to install it on Win10 Enterprise (ver 1607, build 14393.xx ). Any build above this should work I guess?
Bash is running, I haven’t installed Anaconda etc via command-line yet.
I’m totally confused by the Pro/Anniversary/Enterprise names after googling for a day!

I got this error going into Developer mode, this seems to be well documented, and not a dealbreaker.
Developer Mode fails with error code 0x80004005

This was the most detailed description of WSL I saw, valid for more than just the Server version.

Even if someone uses the Ubuntu inbuilt in windows, they might not be able to install PyTorch or will have problem in doing so…
So we are better done with a dual boot or Any cloud computing…

Could you please elaborate?

Hi everyone,

I have a similar topic. I have a pretty powerful rig at home, I want to know if I still need AWS or Paperplate.

-4 NVIDIA GTX 1080Ti with 11GB RAM in SLI
-128GB RAM
-SSD + 2 HDD

Anaconda, pytorch, tensorflow (compiled) and cntk are installed.

AWS or my rig?



1 Like

Definitely your rig.

AWS or Paperspace or other cloud solutions are needed by people who don’t have enough juice on their laptops or systems to make a DL program run/‘train’ in a short enough time period.

I think your rig might even be faster than what the current AWS Instances have to offer(not sure about the P3 ones, they might be a little better).


Thanks for the answer!

My rig it is!

1 Like

I guess it’s too powerful

There is no such things as too powerful ;).

What are you running?
Windows 10 with Ubuntu?
Ubuntu alone ?
Something else?

How many bitcoins per month ??:slight_smile: