marcosf63 / deep-learning-gpu-box-build-instructions Goto Github PK

Deep Learning GPU box build instructions

Instructions to build your own research GPU box for deep learning AND get it set up with tensorflow, pytorch and CUDA drivers.

After following these instructions you'll have:

1. A sick GPU box with multuple GPUs set up.
2. A functional OS ready for deep learning.

Note: I've built 4 of these already and they are amazing. Total cost is around $6,000.

Parts

Considerations

1. Temperature

   • If your machine gets too hot, the GPUs will auto-throttle down.

2. Bottlenecks

   • With deep learnning, the biggest bottleneck is not the GPU but the DATA TRANSFER to the GPUs.
   • This is why the Motherboard needs to be fast enough and should have at least 40 PCI lanes.
   • The drives need to be really fast.
   • Use the SSD to feed data directly to model. Use SSHD for long term storage that won't go into the model directly.

3. RAM

   • Have at least as much RAM as you have GPU RAM.

4. CPU

   • Have at least 1 core per GPU.
   • Water cooling can help keep the overall temperature low.

Assembly / Install instructions

I'll add more details later, but in order you should:

1. Install fans.
2. install motherboard.
3. Install power source (but don't screw in yet).
4. Connect all the fans to motherboard.
5. Install drives.
6. Connect drives to motherboard.
7. Connect motherboard to power supply.
8. Connect power button, usb, etc... to motherboard.
9. Install GPUs.
10. Connect GPUs to power supply.
11. Install RAM.
12. Screw in powersource.
13. Fix all the cables neatly.
14. Insert an ubuntu live USB.
15. Turn on.
16. Pray.
17. Boot into BIOS and set the USB as the priority drive.
18. Install Ubuntu (or your OS).
19. Follow these instructions to set up your system with tensorflow, pytorch and cuda drivers.
20. Learn deeply.

Some deep learning tips

1. Log your experiments and parallelize hyperparameter search using the python library test tube.