3 months ago
- The author built a separate machine named Poppy for local LLM training to avoid tying up their daily PC, Perry, which has an RTX 3090.
- Poppy was originally a small form-factor PC built in 2020 for portability and gaming, with a GTX 1660 Super, but became unused after moving to Lisbon.
- To repurpose Poppy for training, the author upgraded to a larger case (Fractal Design North XL) and a 1600W PSU (ASRock Phantom Gaming PG-1600G) to potentially handle multiple GPUs.
- Initial testing involved training a reduced GPT-2 style LLM on the GTX 1660 Super, which took over 11 days and used more electricity compared to Perry's faster run.
- An RTX 3090 was purchased from eBay to improve performance, but installation caused a CPU overheating issue due to a faulty fan, which was replaced.
- After fixing the fan, Poppy successfully ran a standard LLM training run in about 40 hours, matching Perry's performance, with GPU temperatures stable at 70°C.
- Future plans include water-cooling for the GPU and expanding to multi-GPU setup with a new motherboard and CPU.