I tried warning Apple Silicon LLM adopters. You can buy a $500 GPU with memory bandwidth that is comparable to a $50,000 Mac Studio (and likely faster at GPU compute too).
If looped transformers on smaller LLMs is the future, then the VRAMlets win the war and the Unified Memory crowd overpaid for DDR5.
It's funny, you think you have enough RAM then you realize that what you need is more memory bandwidth (ugh).