Cheap ML Environment

📍 Cambridge, England

It's certainly true that machine learning model training is compute intensive, but can you build a training computer on a budget?

The old gaming PC

Quite a few years ago I had a gaming PC. It's pretty geriatric by today's standards: an old Intel i5 processor and 16GB of RAM, paired to an Nvidia GTX1060 with 6GB of VRAM. It sits under my desk.

I find that it's "reasonably good". I tried transfer learning on my laptop, and one overnight run took seven hours. On this, it takes 50 minutes. Not blazing fast, but "fast enough".

Software

It runs Ubuntu server, which works well enough for what I want from it. It can handle CUDA drivers well enough.

joefg@ml:~ $ fastfetch
                             ....              joefg@ml
              .',:clooo:  .:looooo:.           --------
           .;looooooooc  .oooooooooo'          OS: Ubuntu 24.04.3 LTS (Noble Numbat) x86_64
        .;looooool:,''.  :ooooooooooc          Host: MS-7758 (1.0)
       ;looool;.         'oooooooooo,          Kernel: Linux 6.8.0-88-generic
      ;clool'             .cooooooc.  ,,       Uptime: 1 hour, 44 mins
         ...                ......  .:oo,      Packages: 869 (dpkg)
  .;clol:,.                        .loooo'     Shell: bash 5.2.21
 :ooooooooo,                        'ooool     Terminal: /dev/pts/4
'ooooooooooo.                        loooo.    CPU: Intel(R) Core(TM) i5-3570K (4) @ 3.80 GHz
'ooooooooool                         coooo.    GPU: NVIDIA GeForce GTX 1060 6GB [Discrete]
 ,loooooooc.                        .loooo.    Memory: 812.35 MiB / 15.57 GiB (5%)
   .,;;;'.                          ;ooooc     Swap: 0 B / 4.00 GiB (0%)
       ...                         ,ooool.     Disk (/): 37.38 GiB / 97.87 GiB (38%) - ext4
    .cooooc.              ..',,'.  .cooo.      Local IP (enp3s0): <snip>
      ;ooooo:.           ;oooooooc.  :l.       Locale: en_GB.UTF-8
       .coooooc,..      coooooooooo.
         .:ooooooolc:. .ooooooooooo'
           .':loooooo;  ,oooooooooc
               ..';::c'  .;loooo:'

My development environment on it is purely Python. I do my experiments in a Jupyter notebook, with training details being published to TensorBoard.

If I want to access it when I'm not on my home network, I can do so with Tailscale, which works really well. If I'm out and about with my laptop and I want to check on a training run, I can connect to the VPN and look at TensorBoard.

I also added a notification system that pings messages to my private Telegram bot, so I can see what it's doing when I'm out in the field.

Caveats

  • This old graphics card isn't supported by newer versions of PyTorch, so I'm stuck with PyTorch version 2.6.0 and Python 3.9.

  • It's a bit loud.

  • It's not the most power efficient thing in the world.

  • A powercut means I have to ask someone to reboot the machine.

Back to top