Definition
Unified memory is a single pool of RAM that a chip's CPU, GPU and neural engine all share, as on Apple silicon. Unified memory bandwidth is how many bytes per second the chip can read from that pool. For running language models locally, it is usually the number that matters most after capacity.
Why Bandwidth Sets Token Speed
To generate each token, a model has to read its active weights from memory. If a model's active weights take 20 GB at your chosen precision, and your chip reads 400 GB per second, the ceiling is about 20 tokens per second before any other overhead. Real speeds come in lower. That's why two machines with the same RAM can differ widely in speed, and why active parameters matter: a mixture-of-experts model reads fewer bytes per token than a dense model of the same total size.
Capacity and bandwidth do different jobs. Capacity decides whether a model fits at all. Bandwidth decides how fast it runs once it does.
Apple Silicon Figures, October 2026
| Chip | Max unified memory | Bandwidth | Where |
|---|---|---|---|
| M6 | 32 GB | 170 GB/s | 14-inch MacBook Pro, iMac (October 27, 2026) |
| M5 Pro | 64 GB | 307 GB/s | PROLED MacBook Pro |
| M5 Max, 32-core GPU | 36 GB | 460 GB/s | PROLED MacBook Pro |
| M5 Max, 40-core GPU | 128 GB | 614 GB/s | PROLED MacBook Pro |
| M5 Ultra | 512 GB | 1.2 TB/s | Mac Studio |
Figures are from Apple's newsroom posts and tech specs pages. Note that the lower M5 Max configuration tops out at 36 GB, less than an M5 Pro can take, so the cheaper Max is a poor local-LLM buy.
What to Do With It
Pick capacity first: the model plus its context cache must fit with room to spare. Then compare bandwidth among the machines that pass. Measured results are better than estimates, so check tokens per second benchmarks before you buy, and see MacBook Pro M5 Max vs Mac Studio M5 Ultra for a worked comparison.