All users have the final say! Windows 11's new feature revealed: manual allocation of unified memory

Microsoft is testing a new feature called IntelligentCarveout in an experimental version of Windows 11, allowing users to manually control how the unified memory pool is allocated between the system, graphics card, and AI tasks. It is likely designed specifically for NVIDIA's upcoming RTX Spark chip.

PCs without a dedicated graphics card already need to share the same memory pool between graphics, AI, and system tasks, and this allocation is currently handled automatically by Windows.

As Apple, Qualcomm, NVIDIA, and other vendors build AI PCs around unified memory, with memory pools potentially reaching hundreds of gigabytes, the need for manual allocation is becoming increasingly urgent.

In Windows 11 build 29648.1000, released on August 17, strings related to "IntelligentCarveout" were found, referencing descriptions of reserving memory for accelerators, graphics cards, and AI.

The feature allows users to manually allocate unified memory for gaming and AI applications, with the reserved portion being invisible to other programs. Similar references have also appeared in NVIDIA's documentation.

PCs designed specifically for on-device AI are now shipping with increasingly large unified memory pools. Apple's Macs with 24GB or more of unified memory were once in short supply during the AI boom, and Microsoft and NVIDIA have responded to this trend with RTX Spark devices, offering up to 128GB of unified memory.

If IntelligentCarveout launches officially, users could allocate most of the RTX Spark's 128GB memory pool to graphics and AI tasks. This means the RTX Spark's GPU-accessible memory would far exceed the 32GB GDDR7 on the RTX 5090. However, since unified memory is based on LPDDR5x, its bandwidth is far lower than GDDR7, so this is more of a capacity advantage than a speed advantage.