New Windows 11 setting could give unified-memory PCs more VRAM

Microsoft appears to be preparing a new Windows 11 setting that lets users control how much unified memory the operating system reserves for graphics and AI workloads. The feature could be useful on upcoming PCs that share large pools of memory between the CPU, GPU and other accelerators. Windows Latest found references in Windows 11 Insider Experimental Build 29648.1000 released on August 17. It contains a hidden feature called “IntelligentCarveout”, alongside a new SettingsHandlers_UnifiedMemory.dll component.

“Reserved memory for accelerators” and “Memory for graphics and AI acceleration” were some of the strings found buried within the OS. Apparently, Windows can now reserve additional unified memory for graphics and memory-intensive games and applications.

That makes the proposed feature different from the Shared GPU Memory figure already visible in Windows Task Manager. On conventional PCs, shared GPU memory represents system RAM that the GPU can access when needed rather than memory that is permanently set aside. IntelligentCarveout appears to create an actual reservation by walling off sections of the memory from other applications.

Platforms like AMD’s Strix Halo-based Ryzen AI Max+ 395 can allocate up to 96 GB of its 128 GB unified-memory pool as VRAM. Combined with its 256-bit LPDDR5X-8000 RAM, that provides up to 256 GB/s of memory bandwidth, putting the bus width firmly in mid-range dGPU territory.  While a dedicated GPU with GDDR7 modules will generally be better, buying one in 2026 may not be pocket-friendly.

Nvidia’s upcoming RTX Spark platform might be the reason Microsoft is taking this seriously because it is also designed around a large memory pool shared between its Grace CPU and Blackwell GPU. Windows 11’s memory-management and scheduling behaviour is already being optimised for such hardware, with the company introducing Workload Profile Scheduling for RTX Spark and making changes intended to let GPUs access larger portions of system memory, particularly for demanding local AI workloads.

AMD’s Gorgon Halo platform with its 192 GB of RAM can also benefit from the extra VRAM allocation, as will its rumoured Medusa Halo-based successor. Intel’s Razor Lake-AX is also expected to join the party, albeit at a much later date.  Then again, IntelligentCarveout hasn't reached stable Windows 11 builds yet, and it may stay in Insider builds for the foreseeable future.

In the end, the CPU and GPU still share the same physical memory pool and bandwidth rather than having independently provisioned system RAM and VRAM, and ultimately, that will be the primary bottleneck.