FermiAI — tiny neural networks on a 2011 Fermi GPU

FermiAI is a playground of tiny neural networks that actually run inference on a retro NVIDIA GTX 550 Ti (Fermi, 2011, OpenCL 1.1, <1 GB VRAM) — trained on a modern RTX 5060 Ti and served from the Fermi to your browser.

Concept: Fermi trains & thinks, GitHub distributes. The Fermi is the "brain" (it runs the models); this repo is the "body" (code, weights, and a web UI that talks to it).

What's inside

Toy	What it is	How it runs
LLM Chat	6.98M-param word-level transformer, trained on UltraChat	OpenCL on Fermi; has a built-in calculator (2+2 → 4)
Number Oracle	tiny transformer trained on number sequences	OpenCL on Fermi (3 6 9 12 → 15 18 21 24)
Baby Lab	a / gpu / rare — proof a Fermi can learn arbitrary tiny languages	OpenCL on Fermi
Digit Eye	MNIST digit MLP	OpenCL on Fermi

All models share one C/OpenCL engine (tinyllm.c) — a from-scratch transformer (matmul + LayerNorm + GQA attention + GeGLU FFN) compiled for the Fermi.

Live demo

Open the GitHub Pages site (link in repo About). It talks to the Fermi over the network; the Fermi must be online for the live demo to answer.

Run it yourself (local)

1. Train (needs a CUDA GPU + PyTorch):

python3 train_word.py      # -> tinyllm.bin   (the chatbot)
python3 train_number.py    # -> tinyllm_num.bin (number oracle)
python3 train_a.py         # -> tinyllm_a.bin
python3 train_2word.py     # -> tinyllm_2w.bin
python3 train_rare.py      # -> tinyllm_rare.bin

2. Build & run the engine on an OpenCL device (e.g. the Fermi):

gcc -O2 tinyllm.c -o tinyllm -lOpenCL -lm
./tinyllm --bin tinyllm.bin "hello"        # one-shot reply
./tinyllm --bin tinyllm.bin --server 9001  # TCP server on :9001

Binary format (magic 0x4C4C4D31, ver 2): <magic><ver><vocab><D><NH><NL><BLOCK><FFN_MULT> then per-token utf8 bytes, then float32 weights.

3. Web UI (runs on any PC, proxies to the Fermi):

cd playground
python3 ui_server.py        # http://localhost:8090

Point BACKEND in playground/index.html at your Fermi host (defaults to http://localhost:8090 when served by ui_server.py).

Files

• tinyllm.c — OpenCL transformer engine + TCP server + calculator
• train_*.py — trainers (PyTorch, from scratch)
• word_ref.py, numpy_ref.py — reference validators
• cam_client.py — webcam hand-tracking client that talks to the Fermi
• playground/ — the web UI (HTML + Python proxy)
• draw_diagram*.py — the FermiAI architecture diagrams

License

MIT — do whatever you want with a 2011 GPU.