Memory-bandwidth-bound decode

LLM token generation is limited by how fast model weights stream from memory rather than by raw compute, so high-bandwidth unified memory is what lets large models run locally.

grounded in: Trend: 'Qwen3.5-122B on Mac Studio' and 'large models running on personal hardware' — the enabling fact is Apple Silicon's high memory bandwidth, an established inference property.

Connected concepts

Unified memory inference, CPU inference serving, Quantization, Model offloading, Inference economics