Just finished building this bad boy
This homebrew AI supercomputer can train 10B-parameter models from scratch.
Deep Dive
A Redditor has built a formidable AI training rig using six Gigabyte RTX 3090 GPUs, achieving 144GB of total VRAM. The system features a custom Epyc 7502 CPU setup, modified Tinygrad Nvidia drivers with peer-to-peer enabled, and intra-GPU bandwidth tested at 24.5 GB/s. The builder states it will be used to experiment with training diffusion models up to 10 billion parameters completely from scratch.
Why It Matters
It demonstrates how powerful, pro-level AI training is becoming accessible to dedicated enthusiasts outside big labs.