Windows 11 Tests User Control Over Unified Memory Split for AI Workloads and Gaming

Microsoft Tests Manual Memory Control in Windows 11

Microsoft is experimenting with a Windows 11 feature that would let users manually control how unified memory is divided between system tasks, graphics acceleration, and AI workloads1

1

3

3

. The functionality, internally called IntelligentCarveout, appeared in Windows 11 build 29648.1000 released on August 171

1

4

4

. Code strings reference "reserved memory for accelerators" and "memory for graphics and AI acceleration," confirmed by Windows watcher @phantomofearth3

3

. The feature remains hidden and unavailable through the normal Settings interface4

4

.

Why User-Controlled Memory Reservations Matter for AI PCs

Unified memory refers to a shared pool accessible by the CPU, GPU, and other processors, rather than dedicated VRAM found on discrete graphics cards3

3

. Windows currently manages that pool through dynamic memory allocation, shifting resources between workloads automatically3

3

. That approach works for everyday multitasking but becomes problematic when demanding games and local AI inference compete for the same system RAM3

3

. IntelligentCarveout would let users reserve a portion of unified memory specifically for graphics and AI tasks, keeping it off-limits to background processes1

1

3

3

. A reserved pool could protect GPU memory from competing applications, improving reliability for models that need a fixed memory budget2

2

.

NVIDIA RTX Spark Drives Timing of Windows 11 Changes

The timing aligns with NVIDIA's upcoming RTX Spark platform, which pairs an Arm CPU with Blackwell graphics and supports up to 128GB of unified memory3

3

5

5

. Microsoft confirmed in May that it had already adjusted Windows to allow RTX Spark's GPU to access more of that shared pool3

3

. With Windows 11 developer drivers now available, various notebooks featuring the RTX Spark are expected before year-end5

5

. NVIDIA mentioned during its announcement that running 120B AI models was possible on RTX Spark, hinting that a minimum of 8GB RAM will be reserved as system memory5

5

. The ability to allocate memory for the GPU becomes critical when a 70B or larger model can require tens of gigabytes of addressable memory4

4

.

Performance Trade-Offs and Capacity Advantages

Because RTX Spark's shared memory uses LPDDR5X rather than the faster GDDR7 found on dedicated GPUs like the RTX 5090's 32GB, the advantage is primarily capacity rather than raw bandwidth1

1

3

3

. In theory, RTX Spark's GPU could access far more memory than even the RTX 5090, though that shared pool is considerably slower1

1

. Reservation is not free capacity—allocating 32GB to the GPU on a 64GB system would leave less memory available to Windows applications2

2

. It also does not increase physical bandwidth or turn system LPDDR into GDDR, as performance still depends on the memory controller, GPU architecture, and power limits2

2

.

AMD Already Offers Similar Functionality Through Drivers

AMD already offers something similar through its Adrenalin drivers with Variable Graphics Memory, which lets AMD Ryzen AI Max users manually adjust how much unified memory is allocated to the integrated GPU3

3

5

5

. A 128GB Ryzen AI Max+ 395 system can reserve up to 96GB for graphics through this driver-level control3

3

5

5

. The Windows 11 implementation would represent a significant shift because Windows historically hides most UMA framebuffer allocation behind firmware and driver logic4

4

. Giving end users an OS-level memory-carveout control would mark a notable change as PCs increasingly blur the distinction between system RAM and VRAM4

4

.

What Remains Unknown About IntelligentCarveout

Microsoft has not confirmed the interface, supported hardware, or whether memory can be adjusted without rebooting2

2

. Reports suggest it could eventually live under Settings > System > Advanced, possibly as a slider or preset profiles3

3

. It remains unclear whether the mechanism will coexist with vendor firmware settings such as UMA frame-buffer allocation2

2

. The feature is confined to Microsoft's Experimental Future Platforms channel, a testing branch used well before changes reach general Windows builds3

3

4

4

. Features in this channel can be substantially changed or removed altogether before reaching retail Windows releases4

4

. The concept is particularly relevant to Windows-on-Arm systems and devices with large unified-memory configurations2

2

. Users should not install experimental builds on production systems merely to obtain the feature2

2

.