AFAIK, memory bandwidth. M2 Ultra 800GB/s, whereas M4 Max is just 546GB/s. For example, local LLM inference has a big bottleneck on bandwidth. 50% extra is sign

AFAIK, memory bandwidth. M2 Ultra 800GB/s, whereas M4 Max is just 546GB/s. For example, local LLM inference has a big bottleneck on bandwidth. 50% extra is significant. I wish the Studio received an upgrade, with a new M4 Ultra potentially going over 1TB/s. It also offers better cooling for long computations.