Build your deep learning box: wiki thread

cool setup!
how are the temperatures in summer (exterior. and the PC) ?

Usually the same as exterior temperature, regardless of utilization not more than 30 C.

Hi All, I got a lot of tips about building my own server in this forum. So thank you to all for sharing. I wrote a blog on my experience with building my deep learning server (remote access). Hope those of you starting on this find it helpful.

6 Likes

I am considering Threadripper but the more I read things in this thread, the more confused I get :sob:

I use this site to compare cpu’s.

https://www.cpubenchmark.net

For me in the choice of cpu for my box i compared price vs single thread performace (some of the non ml work i do is single trheaded), number of cores, and overall benchmark vs price. I went intel i7 (lga2066), based on these plus potential to get an i9 in a couple of years without changing the mb.

If you find a good deal on a threadripper and motheboard combo that beats an intel combo id go for it. Saying that ive never used amd, but have read plently of posts from peole who have

Heres a blog about a threadripper dl rig build.

1 Like

A lot of libs are less tuned for AMD than for Intel. So even if your AMD processor is in theory faster, it may not work that way in practice. E.g. anaconda comes with the MKL linear algebra library that’s written by intel and tuned for their CPUs.

1 Like

Ha! Too late. I ordered an AMD Ryzen 7 1800X last night. I think it’ll be mostly OK though, 8 cores at decent speed

1 Like

It’ll be fine - just make sure that you use AMD optimized BLAS libs. I haven’t used AMD for a while so I’m a bit out of date - hopefully they’re included with anaconda nowadays…

2 Likes

1080ti is getting restocked today…700$ …Get your hands on it asap :slight_smile:

2 Likes

Where?

got an email

1 Like

Very interesting thread.

One note: power supply is important. Not only power of it, but also quality. I had an issue with my DL box crashing and restarting when training models. The issue was narrowed down to the power supply. https://github.com/tensorflow/tensorflow/issues/8858

I ended up buying this one: Cooler Master V1000

1 Like

EVGA site has a way to measure the power supply needed based on the size and capacity of your computer.

https://www.evga.com/power-meter/

3 Likes

PCPartpicker will also add up the total power needed for your build.

Thanks for the link!
My point was that even if you choose power supply based on the specs of your PC, you can still have issues. My DL box was restarting randomly despite the fact that I had 750W power supply that was more than enough to support one GPU.

I started considering AMD since the recent meltdown/spectre issues, and kept looking for clues if AMD can replace Intel for deeplearning.

This :point_up_2: clears all my doubt, thank you!

TL;RD - sticks with Intel (as of March 2018).

Nice article! Did you get these in India?

Also for the last bit you mentioned there - use ngrok? It’s free for one instance running. You can run it on same port as jupyter and keep using same ngrok domain as long as you don’t shut down the ngrok server (the HDD swap is gonna come in handy when it hibernates). Don’t forget to add some authentication to jupyter btw. :slight_smile:

1 Like

Hi Bharadwaj, I got the parts in US (newegg / amazon). I setup remote access I used similar to what is mentioned in this link without additional software:

I have not tried ngrok. Seems interesting.

1 Like

According to the Task Manager, the GPU1 (GTX1070) is not in use at all.

There is no error message if torch.cuda.set_device(0). However, I got error message if set_device(1). Am I missing anything in setup? My OS is Windows 10.

You only have 1 nvidia GPU, so you can only set_device(0).

2 Likes