Unified memory explained
Unified memory is a design in which the CPU, GPU and neural accelerators share a single pool of RAM instead of the GPU having its own dedicated VRAM. For local LLMs this matters because model weights must sit in memory the GPU can reach. On a unified system, most of the total RAM can hold the model. Apple Silicon Macs use this design, and in this catalog they range from 32 GB at 170 GB/s (Apple Mac mini (M6)) to 128 GB at 614 GB/s (Apple Mac Studio (M5 Max)). AMD's Ryzen AI Halo uses a similar shared 128 GB pool at 256 GB/s. The trade-off is that unified memory is soldered: you choose capacity when you buy and cannot upgrade it later. Bandwidth, measured in GB/s, determines how quickly tokens are generated.