Apple's website says the new M5 Ultra chip provides up to 512GB of
high‑bandwidth unified memory and up to 1.2TB/s unified memory bandwidth, about
50% higher than M3 Ultra, enabling very large datasets to be kept in local
memory, higher tokens/sec and on‑device execution of LLMs with hundreds of
billions of parameters. Apple says the M6 supports up to 32GB of unified memory
to run multiple high‑load apps and on‑device LLMs for privacy‑focused AI agent
tasks, and offers up to 170GB/s unified memory bandwidth, roughly 10% above the
M5 and about 2.5x the M1.
