Unified memory (Apple Silicon)

In short: On Apple chips, the processor and graphics unit share one memory. This lets the GPU use a large part of the total memory.

In more detail: A Mac with 128 GB of memory can load models unthinkable on single cards. Bandwidth is 100 to over 800 GB/s depending on the chip, so faster than normal PC memory but slower than high-end graphics cards.

Our context: The AI speed calculator includes many Apple chips (M1 to M4, Pro, Max, Ultra) with bandwidth and selectable memory.

In Depth

• How much is usable: By default about two thirds to three quarters of memory for the GPU, adjustable.
• Software: MLX, llama.cpp with Metal and Ollama run well.
• Limits: Prompt processing (prefill) is slower than on large NVIDIA cards because compute is lower.
• Alternative from AMD: “Strix Halo” systems also use shared memory.

See also: Memory bandwidth, VRAM and model size