Skip to main content

​
Compute

Our advice is to go towards building a distributed compute system (homelab) that consists of a desktop and a laptop. The desktop can be left on 24/7 and use your existing laptop with Tailscale to remote to it. In what follows, the bulk of the cost is on the always on desktop and the laptop is thin and light aka. inexpensive. Our recommendation below is based on the situation in the market for 2026, which is dismal due to memory and storage shortages.

​
Desktop

  1. Good option for up tp $2000 cost: If you already have a desktop and plan to buy a dedicated GPU, choose NVIDIA, CUDA remains the standard for AI workload acceleration. If you already have an AMD or Intel GPU, it may still work for some tasks. Choose memory over raw compute - the memory should be at least 16GB VRAM.
  2. Good option for up to $4000 cost: If you dont have a desktop, instead of getting a desktop that has a GPU limited to the typical 16-32 GB of VRAM, I advise to get the Nvidia spark. Despite its limitations in training, most of the time you will do fine tuning. Its 128gb of VRAM is bulletproof for at least three years. Another option in this category is the compact desktops such as the MINISFORUM MS-S1 Max Desktop Workstation or similar that have the specific AMD chip with unified memory of 128GB of VRAM.
  3. Good option for up to $4000 cost: If you trust the Mac supply chain and the support you get, you can also get the Mac mini or Mac Studio with at least 64GB of unified RAM. Always prioritize unified memory (RAM) over storage.

​
Laptop

  1. Good option for up to $1000 cost: Any reliable non-mac laptop that has at least 16GB of RAM. You will use it for doing remote docker development into the desktop using VS code and Tailscale. Prefer used enterprise laptops rather than cheap consumer-grade new laptops.
  2. Good option for up to $2000 cost. If you insist on the mac ecosystem, I really advise not to buy the most expensive laptop in the store but invest on this hybrid desktop-laptop solution. Get a Macbook pro in this price range.

​
Storage

Do not over invest in laptop or desktop nvme storage - instead get a network attached storage device (NAS) and use 8/12/16TB spinning discs until the solid state storage crisis is over. Then you can augment the spinning disks with nvme cache drives. You cant go wrong with this NAS or common hardware that supports the TrueNAS OS.

​
Google Colaboratory

Google Colab provides free CPU/GPU resources. Log in with your university Gmail account to access expanded Google Drive storage. Colab is useful for demonstrating that your results are replicable and for notebooks that require Colab-specific features. Even when using Colab, you must submit work via GitHub.

​
Cloud GPU

There is a number of inexpensive cloud provides to choose from such as Modal and Hugging Face. HF Spaces offers free Docker-based environments which is good option when assignments or projects require container-based development without local GPU. HF also offers GPU environments for a price.

​
Virtualization

We use Docker containers for all course work. The guide below covers installation for Windows, macOS (including Apple Silicon M-series), and Linux, plus GPU acceleration options.

Docker Containers

Docker installation, course environment setup, Apple Silicon / MPS acceleration, ROS on Mac, and all build/run commands.

​
Managing Python dependencies

Python & uv

Installing Python, managing virtual environments, and installing packages with uv, the fast, modern replacement for pip and conda.
For Git setup and GitHub workflow instructions, see the Git & GitHub guide.
Edit this page on GitHub or file an issue.