Unified memory

Unified memory is a design in which the CPU and GPU share one pool of memory instead of the GPU keeping dedicated VRAM. Both processors address the same memory, so data need not move between system RAM and a graphics card. CUDA's Unified Memory is a separate software feature. It lets CPU and GPU code share managed allocations on every CUDA system, usually migrating pages when the GPU has its own memory. On Grace Hopper and Grace Blackwell hardware, the coherent NVLink-C2C link lets the CPU read GPU-resident managed data without migration.

NVIDIA's DGX Spark shares up to 128 GB of LPDDR5X between the CPU and GPU of its GB10 Superchip, over a 256-bit interface at 273 GB/s. NVIDIA rates the 128 GB version for models up to 200 billion parameters. An RTX PRO 6000 Blackwell card holds 96 GB of GDDR7 on a 512-bit bus at 1,792 GB/s. The Spark fits a third more data, while the discrete card reads its memory about 6.6 times faster.

Choose unified memory when a model or dataset will not fit in a candidate card's VRAM and fitting it matters more than speed. Choose a discrete GPU when the working set fits in its VRAM and bandwidth sets the pace.

Sources

Last verified October 7, 2026.