Unified Memory

Unified memory is a memory architecture where CPU and GPU share the same physical RAM pool, eliminating CPU↔GPU copies. Apple Silicon and AMD Strix Halo (Ryzen AI Max) use this; modern NVIDIA Grace-Hopper and GB10 systems also expose unified memory across chips.

For local AI, unified memory is the reason a 128 GB M3 Ultra Mac can load and run a 120 GB model without dedicated VRAM — something no consumer NVIDIA card can do.

The tradeoff is bandwidth: unified DDR/LPDDR (200–600 GB/s) sits between consumer GDDR6X and HBM. A model that fits on both an M-series Mac and an RTX 4090 will usually run faster on the 4090 due to bandwidth, even though the Mac has more capacity headroom.

An operator comparing an M3 Ultra Mac Studio (192 GB) against an RTX 4090 (24 GB VRAM) for running Llama 3.1 70B faces a capacity-versus-bandwidth tradeoff directly caused by unified memory. The 4090 physically cannot hold a 70B model even at 4-bit quantization (roughly 40 GB) without offloading layers to system RAM over PCIe, tanking throughput. The Mac Studio loads the full quantized model into its unified pool in one shot, no offload needed, because CPU and GPU draw from the same 800 GB/s LPDDR5X pool. But for a 7-8B model that fits in 24 GB VRAM outright, the 4090's GDDR6X bandwidth (1008 GB/s) plus dramatically higher compute throughput wins on tokens/sec — unified memory buys capacity headroom, not raw speed. Check /will-it-run before assuming either platform wins by default.