In the world of personal computers, the boundary between main memory (RAM) and video memory (VRAM) has long been treated as inviolable. The CPU uses DRAM on the motherboard, while the graphics card occupies dedicated high-speed VRAM on its own board. "Unified memory architecture," which breaks down this physical separation and allows processors to share the same memory space, became widely known through Apple's M-series chips.
In the Windows environment too, a shift in design philosophy is beginning with the arrival of new-generation silicon featuring large-capacity unified memory, such as NVIDIA's RTX Spark and AMD's Ryzen AI Halo. In mid-August 2026, a hidden settings screen for freely allocating unified memory specifically for graphics and AI processing was discovered within the internal structure of an experimental Windows 11 build. A new control layer is poised to be added to Windows' memory management model, which had remained fixed for many years.
| Configuration | CPU-GPU Memory Relationship | Data Handling |
|---|---|---|
| Discrete memory configuration | CPU uses system RAM, GPU uses dedicated VRAM | Depending on the processing involved, data transfer between separate memory regions may be required |
| Unified memory configuration | CPU, GPU, and NPU share a common memory pool | May reduce unnecessary copying, but actual allocation and bandwidth depend on hardware and OS implementation |
"IntelligentCarveout," an Undisclosed Feature Found Deep Inside a Test Build
The community that tracks Windows' internal structure was the one that caught this sign. A settings handler not present in any stable release was found in the system binaries of Windows 11 build 29648.1000, distributed to the Feature Platforms Experimental channel on August 17, 2026. Its existence was revealed through reporting by Windows Latest and verification by community watcher @phantomofearth.
A new library file named SettingsHandlers_UnifiedMemory.dll had been added within the build, and a screenshot of it was shared by @XenoPanther. This internally-named feature, "IntelligentCarveout," has been assigned feature ID "61121285." Strings within the code directly embedded descriptions such as "Reserved memory for accelerators" and "Memory for graphics and AI acceleration."
However, Microsoft has not disclosed this feature in official release notes or elsewhere. While it is possible to forcibly enable the flag using the third-party tool ViveTool to display the settings item, the code remains at a development stage that is disabled by default even for Insider participants. Because using unofficial tools carries a risk of undermining system stability, enabling it in a general-use environment is not recommended.
The Full Picture of a UI That Carves Out "Accelerator-Reserved Space" via Sliders and Presets
The settings interface that was actually invoked has a concrete structure for directly controlling hardware resource allocation. Its on-screen location sits within "Settings > System > Advanced settings," appearing as an item called "Memory for graphics and AI acceleration" inside a newly established "Unified memory" section.
The description accompanying the settings screen explicitly states: "Allows Windows to reserve additional unified memory for games and applications that heavily use graphics and AI. Reserved memory cannot be used by other applications." This suggests a design philosophy of setting aside a specified amount of capacity for accelerators, which differs from the dynamic memory sharing that Windows has traditionally performed.
| Item Checked | Content Confirmed in the Experimental Settings Screen |
|---|---|
| Settings location | Settings → System → Advanced settings → Unified memory |
| Item name | Memory for graphics and AI acceleration |
| Options | Custom, Recommended, High, Maximum, Don't allow |
| Adjustment method | A manual adjustment slider is provided in addition to presets |
| Caveat | This is an unfinished UI invoked from a test build; the names and options are not guaranteed to be adopted as-is in the final product |
An important limitation is that this settings item does not function on every Windows 11 device. A hardware gate is in place so that it only appears on unified-memory devices in the Copilot+ PC class equipped with an NPU (Neural Processing Unit). On a typical desktop environment with a discrete GPU plugged into the motherboard, this item does not appear at all. Since this is an experimental feature in an early stage of development, there remains room for the UI design and behavior to change in future OS updates, or for the code to be removed entirely without ever reaching official adoption.
The Line Between Microsoft's Already-Announced "Optimization Plan" and the Hidden Feature
While the UI slider found deep within the build is an unannounced discovery, the fact that Microsoft is planning fundamental OS-level revisions for unified memory environments is already public knowledge. On May 31, 2026, during NVIDIA GTC Taipei held in Taiwan, Microsoft officially announced its policy for unified memory optimization aimed at RTX Spark via the Windows Experience Blog.
That announcement declared that Microsoft would "set a new, smarter upper limit on the total amount of system memory the GPU can access." This was explained as enabling the deployment of large language models (LLMs) running locally on PCs, as well as the execution of extremely complex real-time rendering processes. Microsoft further revealed that it was strengthening the internal mechanism for expanding the page size managed by Windows in shared memory regions on unified memory systems, allocating huge memory pages for heavy workloads.
| Category | Confirmed Content | Current Status |
|---|---|---|
| Memory management improvements officially announced by Microsoft | Expansion/optimization of the memory ceiling available to the GPU, and improved management of shared memory page sizes | Officially announced. However, specifics regarding the scope of provision and implementation details remain to be revealed |
| IntelligentCarveout confirmed in build 29648 | A settings screen for specifying, via presets or a slider, the amount of memory capacity for graphics or AI processing | An unannounced experimental feature. Inclusion in the final product is not guaranteed |
As the tech outlet TechPowerUp pointed out, Microsoft also hinted, during its Computex announcement, at providing developers with an environment that flexibly optimizes memory workloads between the CPU and GPU. However, what was discussed at the time was an automatic memory space management method handled by the OS kernel or driver layer, not a control screen like IntelligentCarveout, where end users manually determine allocation in megabyte increments by moving a slider.
The officially announced "automatic internal OS memory expansion and page size optimization" and the now-discovered "user-driven memory reservation UI" should be clearly distinguished from one another, even as they exist in a complementary relationship.
New-Generation Hardware Is Driving OS-Level Memory Restructuring
Why is this kind of unified memory control needed in Windows right now? The background lies in the structural shift of new-generation silicon set to hit the market from late 2026 onward.
Leading the charge is NVIDIA's "RTX Spark," announced on May 31, 2026. It integrates a Blackwell architecture GPU with 6,144 CUDA cores and a 20-core Grace Arm CPU into a single package, paired with up to 128GB of LPDDR5x unified memory. Its AI processing capability for single-precision and low-precision operations reaches a claimed 1 petaflop (FP4). A fall 2026 launch has been announced from companies including Surface, ASUS, Dell, HP, Lenovo, and MSI.
AMD, meanwhile, is accelerating a similar approach through its developer-oriented platform, "Ryzen AI Halo." Built around the Ryzen AI Max+ 395 processor, it is equipped with up to 128GB of LPDDR5x memory running at 8,000 MT/s, achieving 256 GB/s of memory bandwidth. It integrates a Radeon 8060S for graphics and an XDNA 2 architecture NPU for AI processing.
AMD has already gotten ahead of Microsoft by offering "Variable Graphics Memory (VGM)" through its own utility software, which variably allocates unified memory. It established a mechanism allowing users to manually set the capacity equivalent to dedicated VRAM within a range of 10% to 90%, in 5% increments.
| Platform | Maximum Unified Memory Capacity | Memory Specs / Bandwidth | User Memory Allocation Control | Market Rollout Status |
|---|---|---|---|---|
| NVIDIA RTX Spark | Up to 128GB LPDDR5x | High-bandwidth unified interconnect | Planned integration with a new Windows OS-side feature | Various manufacturers' hardware shipping sequentially from fall 2026 |
| AMD Ryzen AI Halo (Ryzen AI Max+ 395) | Up to 128GB LPDDR5x | 8,000 MT/s (256 GB/s) | Variable allocation of 10-90% via proprietary utility (VGM) | Deployed as a developer-oriented platform |
| Apple M-series (for reference comparison) | Up to 128GB-192GB (Max/Ultra configurations) | Unified memory (over 800 GB/s max) | Fully automatic dynamic allocation by the OS (no manual UI) | Widely established in the general market |
Unified memory architecture in PC processors was pioneered in practicality and convenience by Apple's Mac lineup. That is now spreading across the entire Windows ecosystem, on both Arm and x86. With the boundary line on the hardware side having disappeared, a challenge has emerged on the software side as well: how to define the power balance between the CPU and accelerators.
The Trade-offs Between "Shared GPU Memory" and the Pre-Reservation Approach
Windows has long had a framework called "Shared GPU Memory." Any user can check this figure by opening the Performance tab in Task Manager. However, the intended use case differs between the newly discovered IntelligentCarveout and the existing Shared GPU Memory.
Conventional Shared GPU Memory is simply an "upper limit (ceiling) on system RAM that the GPU can borrow when needed." Generally, roughly half of installed memory is allocated as this ceiling, and the graphics driver works in coordination with the Windows memory manager (DMM) to dynamically juggle memory pages. Under normal conditions, it is freely used by ordinary CPU-side processes, and it serves as a temporary overflow space the moment the GPU-side VRAM runs dry.
| Approach | How Memory Is Allocated | Impact on Regular Applications |
|---|---|---|
| Existing Shared GPU Memory | The GPU dynamically uses system memory as needed, within an upper limit | Not a permanently fixed region; usage fluctuates depending on load and available capacity |
| Experimental IntelligentCarveout settings screen | The user specifies in advance the capacity to be used for graphics or AI processing | According to the on-screen description, reserved capacity becomes unavailable to other applications. However, the low-level implementation method and the scope of accelerators covered have not been confirmed |
The intent behind this setting is thought to be making it easier to secure, ahead of execution, the amount of memory available to graphics and AI workloads. While this may allow for more stable capacity, increasing the reserved amount would decrease the memory available to ordinary applications. However, since Microsoft has not published technical documentation for IntelligentCarveout, it would be premature to conclude that the mechanism guarantees physical memory partitioning, bandwidth reservation, or full exclusive occupation by the GPU or NPU.
If it is indeed implemented as described on screen—reserving capacity in a way that makes it unavailable to other applications—it could make it easier to estimate available memory in advance for tasks such as loading large local generative AI models or unpacking vast texture data. That said, effects such as preventing memory fragmentation or guaranteeing throughput have not been confirmed.
In exchange, there could be trade-offs in system resources. For example, if 96GB out of 128GB of unified memory were reserved for AI processing under an implementation where the entire reserved amount becomes unavailable to general applications, only 32GB would remain for browsers, office software, and the OS itself. However, it remains unclear whether there is a mechanism to return reserved space as needed, which games or AI workloads would be granted access, and how Windows would determine priority among applications.
Challenges Ahead of Fall 2026 Hardware Launches
The code found in Windows 11 build 29648 directly illustrates Microsoft's trial-and-error process in responding to the structural transformation of PC hardware.
The hardware-side timeline is clear. Machines equipped with NVIDIA RTX Spark are confirmed to begin shipping in fall 2026, and AMD's Ryzen AI Max platform is also advancing its rollout. Unless the OS side establishes a means of efficiently handling unified memory by the time new-generation PCs with massive unified memory reach consumers' hands, the hardware's potential cannot be unlocked.
However, there is no guarantee that an experimental feature will make it into the final, official Windows product as-is. In Microsoft's preview builds, it is a routine occurrence for features prototyped in early development stages to undergo major spec changes based on feedback and internal testing results, or to be discarded entirely without notice. As Neowin has also reported, caution is warranted against placing excessive expectations on features still under development.
Will the manual settings slider approach survive, or will it be consolidated into the fully automatic, low-level optimization announced in May? Or will it be offered in limited form as a differentiating feature for NPU-equipped Copilot+ PCs? The answer to the question of which—manual or automatic—truly delivers a superior user experience should take on a clearer shape alongside the arrival of new hardware this fall.