LocalLLaMA debates best budget multi-GPU training rig

This is less a product announcement than a reality check on local LLM infrastructure: memory-per-dollar still matters, but interconnect and software compatibility decide whether a cheap card is actually usable.

• –MI50 gets some credit for inference, but commenters say weak Python package support and missing BF16/FP32 training advantages make it a bad default for modern fine-tuning
• –V100 32GB setups still look compelling for budget multi-GPU work, especially when NVLink is available and workloads do not fit on one card
• –RTX Pro 6000 is framed as the cleanest high-end option, but its price moves it out of the practical range for most hobbyist builders
• –The thread treats raw VRAM as only part of the equation; GPU-to-GPU bandwidth and ecosystem maturity are the bigger bottlenecks once training scales past one card