Local AI Homelab Weighs V100, MI50

Datacenter salvage is the fun way to buy VRAM, but it is rarely the clean way to buy progress. In 2026, V100 looks like the safer "works this weekend" pick, while MI50 only makes sense if you actually enjoy ROCm archaeology.

• –NVIDIA still keeps V100 in current data-center docs, and vLLM's CUDA path explicitly supports V100-class GPUs: https://www.nvidia.com/en-us/data-center/v100.md/ https://docs.vllm.ai/en/v0.11.2/getting_started/installation/gpu/
• –AMD's MI50 page shows the 32GB, 300W passive-server reality, and ROCm's issue tracker quotes release notes that put gfx906 cards, including MI50, into maintenance mode starting Q3 2023: https://www.amd.com/en/support/downloads/drivers.html/accelerators/instinct/instinct-mi-series/instinct-mi50-32gb.html https://github.com/ROCm/ROCm/issues/2308
• –My read: if the goal is learning local AI rather than debugging drivers, stay in the consumer-GPU lane for now or save for a newer single card; the cheap legacy-card route is usually a project in itself.