Sjoerd’s Rig

Community rig for deep learning

This deep learning rig to supports the work of me and my community.

It has fast ssd’s and a lot of memory to support extensive data transformations that is found in computer vision and some graph neural networks. It also great for Bayesian machine learning and LLM`s.

The specs are:

• 4x NVIDA GeForce RTX 3090
• 1x NVIDA GeForce RTX 4090
• 2x PCIe-4 SSD’s
• 24 cores AMD Epyc 7402P
• 2x 1800 W PSU
• 512 GB memory
• Arch Linux
• Micromamba for package management
• Docker to run containers
• Fish terminal
• Tmux
• Neovim

Why I build a deep learning rig?

From 2018-2023 I have worked as R&D engineer deep learning at large tech companies.

Usually I worked with about 20 deep learning specialists, all in need of GPU’s.

The GPU cards that we used were about $50 0000 each, and they were often overbooked.

When I started taking on more freelance clients, I wanted to have a similar environment to support creative ideas, but not at the same cost.

It became a sport to build a machine with similar capacity at a fraction of the price.

I was influenced by Jeremy Howard from fast AI, who managed to [train imagenet in 3 hours for $25] in 2018 with his “deep learning study group”.

Jeremy his thesis is “you don`t need the best hardware; creativity is all you need”