Reading up on M3 Ultra

For local LLMs, memory bandwidth and size determine token speed and model capacity; Strix Halo offers more memory per dollar, while Apple Macs with higher bandwidth generate tokens faster.