Description

Built a computer for my Deep Learning projects at Stanford University because AWS and Google Cloud was too expensive. www.thisisjeffchen.com.

[July 2019 Update --> see here for the new parts list with new processor and GPU https://pcpartpicker.com/b/nrjypg]

If you're a grad student working on Machine Learning, ML enthusiast, startup guy in ML, or a ML weekender, you will need GPU resources to train your models for experiments. I frequently run 12 experiments across 4 GPUs, which is makes model development go 12x more quickly. (Start exps, go to sleep, wake up and check results, repeat).

But it's expensive to get a quad GPU machine (each GPU is $700+) and if you don't know how much GPU power you'll need, the best idea is build a computer with 1 GPU and more GPUs as you go along.

12 Core CPU is chosen because I've run 4 experiments / GPU before, so that's 4 threads or 2 cores per GPU. The extra 4 cores are left over so the machine can do other stuff while running experiments.

M.2 SSD is way faster 3.4GB/sec than SATA3 600MB/sec.

1080 Ti's markup over 1070 Ti is worth it since the extra speed and 3GB comes in very handy and can help you go much faster in training (larger batch sizes, more CUDA cores).

DDR4 3000 Quad channel memory chosen as Threadripper has a big performance boost when in quad channel configuration.

1600W P2 power supply chosen for noise considerations and ability to handle 4 GPUs. (G2 is very loud).

S24 cooler chosen for noise considerations and aesthetics (Air cooled options are so ugly and block too much of the motherboard).

Lian-Li PC-O11AIR Case chosen because it has 8 PCI-e slots, great airflow, and looks nicer than the Corsair Air case.

[Notes: 1 - Memory actually shows CMR64GX4M4C3000C15, which is in short supply, so I replaced it with CMW64GX4M4C3000C15, exact same specs but not officially listed as supported by TR.

2 - photos show 4 GPUs, I added them later, which is not reflected in the parts list. One is a 1080 Ti Founder's Edition, one is a Titan RTX, and the other is a Asus 2080 Ti Turbo ]

Comments

  • 13 months ago
  • 2 points

Great setup, thanks for sharing. Did you know that you can apply for academic GPU grants directly from NVIDIA? The approval rate is quite high. Good luck with your projects!

  • 13 months ago
  • 1 point

Good to know, thanks!

  • 12 months ago
  • 1 point

Great Post!

I'm a graduate student looking to upgrade my current computer (see link in post) to run high memory computations. I was curious whether or not you think I should utilize my current CPU (with a new motherboard that supports more RAM) or if I should just rebuild from scratch. I don't do deep learning, but definitely plan to do a lot of bootstrapping and data analysis with large datasets (5-100+ GB). I do have access to our University supercomputing cluster, however I would like to have access to my own computing device as the queues can get quite long and to do downstream data analysis off the cluster. Any advice would be greatly appreciated.

https://pcpartpicker.com/list/Fv88mq

  • 12 months ago
  • 1 point

If you are not doing deep learning, and your workload will be CPU bound. Then it depends on if you can multithread the analysis. If yes, then going to something like a 16-32 core CPU should help. If not, then your current CPU will work fine.

  • 12 months ago
  • 1 point

Howdy! I really like this build but need more RAM, hopefully 256GB. Can you recommend a CPU/motherboard pairing to best facilitate this?

  • 12 months ago
  • 1 point

You will likely have to use a dual socket setup...

  • 12 months ago
  • 1 point

argh. what about just 128GB RAM? I think the CPU you recommended supports that, but will there be space on the motherboard for 8 RAM sticks plus 4GPUs? Planning to put in 4 2080Tis... Thanks!

  • 9 months ago
  • 1 point

yes there should be space. just look at the motherboard

  • 11 months ago
  • 1 point

How important is ecc memory for 1) CPU 2) GPU

  • 10 months ago
  • 1 point

Well these aren't really options for a consumer machine... generally for deep learning because there are millions of parameters a few corrupt ones shouldn't make much of a difference.

  • 11 months ago
  • 1 point

Jeff, thank you for sharing. This is great!

1080 Ti seems out of stock in most sites now. Is 2080 Ti worth the money and does 2080 fit the case Where did you want the parts? Neweggs / Amazon etc?

  • 10 months ago
  • 1 point

I still like 1080 tis... because im not sure 2080 tis are great value. have you tried eBay.

https://medium.com/the-mission/how-to-build-the-perfect-deep-learning-computer-and-save-thousands-of-dollars-9ec3b2eb4ce2

  • 11 months ago
  • 1 point

Really nice case and build, looking to do my own build with it. Does the case support a 120mm x 240mm fan/radiator on the top for a CPU cooler or does it need to be 280mm or 360mm?

Thanks,

  • 10 months ago
  • 1 point

im using a 120x240 on top for mine.

  • 9 months ago
  • 1 point

if you add all 4 gpus, is there space enough on the bottom for fans?

  • 6 months ago
  • 1 point

There will not be any more space on the bottom for fans w/ 4 GPUs. Fans need to be mounted on the front and possibly the front side.

  • 9 months ago
  • 1 point

where can I buy 1080ti with your price?

  • 6 months ago
  • 1 point

I think you need to look on eBay now since it's no longer being made.

  • 7 months ago
  • 1 point

Cool setup! I've read conflicting information regarding using AMD vs Intel chips. I can see that the AMD Threadripper series is more value for money from a cores/threads perspective. But, is there an issue with optimization when using ML libraries? I will be using standard libraries such as TensorFlow, PyTorch, DarkNet, AlexNet etc, largely for computer vision applications (image processing, learning etc). Lastly, any motherboard recommendations with regard to using the i9 9900K that can be future upgradable with 4 GPUs? Many thanks!

  • 6 months ago
  • 1 point

I got the same mobo but wish I got one with built in wifi module in the i/o slot. Now I need to use a PCI slot for the wifi card :(

  • 6 months ago
  • 1 point

So, I'm confused... the mobo on this list actually does have a wifi card built in, no? I'm looking at the version on NewEgg and that is what it seems to say. Did you get some other variant?

  • 5 months ago
  • 1 point

No, the MSI gaming pro carbon x399 doesnt have the wifi card built into the IO shield. It comes as an additional PCIe1 card. Kind of annoying since it will take up one slot. If you want to max out at 4 GPUs, the wifi card gets in the way.

  • 6 months ago
  • 1 point

I am about to use this build spec to assemble a machine, but I wonder what is the tradeoff between (a) A single 2080 Ti, vs (b) A pair of 2070s? The second option gives more cores and tensor cores, total (for less money) ... but does it matter that each of the 2070s has only 8GB memory, I wonder?

  • 5 months ago
  • 1 point

Great!

I saw your post on medium. The price for EVGA SuperNOVA 1600 P2 is 260$ only. But price listed here is 378$. Was it different version there?

Thank you!

  • 3 months ago
  • 1 point

Would you recommend upgrading to the new Ryzen CPU like so? https://pcpartpicker.com/list/JB7Gvn

  • 1 month ago
  • 1 point

Thanks for sharing this great build! I am using this as a model for my own deep learning machine.

What are you thoughts on cheaper motherboards with similar features, e.g. Gigabyte X399 AORUS PRO or Asus PRIME X399-A?