Building deep learning workstation for $2200


Gave the examples/text notebook a go for AWD-LSTM, but the model gave an error when I tried to move it to fp16:
“Expected tensor for argument #1 ‘indices’ to have scalar type Long; but got torch.cuda.HalfTensor instead (while checking arguments for embedding)”
Points to torch/nn/ so appears to be upstream of fastai.

1 Like

(Andrea de Luca) #23

Sound quite related to fp16…

Keep us posted, should you manage to get it working!


(Scott H Hawley) #24

Big relief to read you says “28 PCIE lanes should be ok for three GPUs”.
Have you built your system – how is it?

I’m building a system with an i7-7800X and two RTX 2080Ti’s, but got worried that since the CPU can’t do full 16x twice, that I’d hit a bottleneck. (I’m used to working with Xeon CPUs which have lots of PCIE lanes)


(Andrea de Luca) #25

According to Tim Dettmers, who ran a series of systematic experiments, x8 gen3 does actually provide sufficient bandwidth for 2-3 cards. Note also that a single xeon cannot provide the 48 lanes necessary to run 3 cards at full x16 (skylake xeon-w could, but 4-8 lanes are always used for other stuff).

1 Like

(Scott H Hawley) #26

Thanks, I saw that too (after I posted here)!

Just posted a “Completed Build” on PCPartpicker:

The budget for this one was $4000, so I got extra RAM and a second GPU.


(Scott) #27

Hey, guys
I am not tech savvy and building a PC looks pretty complicated for me as well as lack of tech support in case of issues. Does it worth it to buy a prebuilt deep learning desktop? I’ve found a few companies focused on this.

Any thoughts on this?


(Malcolm McLean) #28

Thoughts are that it depends on your time and resources. What I did (two years ago) was to buy a used gaming computer with GTX 1070 off Craigslist for $800. I see now on several with suitable nvidia GPUs (1070/1080) for $500-$1000.

Because I am a PC and Linux ignoramus, I hired a Linux expert to assess the purchase and get me through a couple of support issues. He works by phone, charges per hour, and was also found via craigslist.

So if you have the time and a little expertise you can trade it for saving money.

Other thoughts:

  • Used last generation nvida GPUs are quite cheap. Miners quitting and gamers upgrading to the latest generation.

  • Cloud computing looks like a very reasonable option right now, and the one I would choose today. You can get started for free with Google Colab, and later upgrade to a faster paid service. $3000 will pay for a lot of hours, plus you are freed from hardware failures, depreciation, and system maintenance. And you can easily work from any location, without the complications of port forwarding, WOL, dynamic DNS, VNC, and SSH.

Hope these comments are more clarifying than complicating!