Windows 11 is getting an option to change how much unified memory is set aside for graphics and AI work, so you can free up more RAM for other applications or reserve more for running a local AI model, depending on what you’re doing.
Unified memory is a single, high-bandwidth memory pool shared between the CPU, GPU, and other accelerators, instead of the usual split between system memory and dedicated video memory. It was popularized by Apple with its M series and is now proving to be a boon for local AI applications.
Windows Latest first spotted references in internal NVIDIA documentation that pointed to Windows gaining more control over how unified memory is reserved for graphics and AI workloads. It turns out that the recent Windows 11 preview builds also contain those references, which was verified by Microsoft watcher phantomofearth to Windows Latest.
In build 29648.1000, released August 17, Windows includes a hidden feature called IntelligentCarveout with the feature ID 61121285, along with a new SettingsHandlers_UnifiedMemory.dll file.
More importantly, Windows 11 also contains the following strings:
“Reserved memory for accelerators”
“Memory for graphics and AI acceleration”
“Let Windows reserve additional unified memory for graphics and AI intensive games and applications. Reserved memory is not available for other applications.”
Of course, as it’s an unreleased feature, Microsoft’s release notes say nothing about this.
Windows 11 is preparing to let you control how unified memory is reserved
A normal laptop with a discrete GPU splits memory into two buckets: system RAM for the CPU and VRAM for the GPU. However, NVIDIA’s RTX Spark platform uses unified memory, with up to 128GB shared between a Blackwell GPU and a Grace Arm CPU. So do AMD’s Ryzen AI Max “Halo” series and other upcoming platforms that rely heavily on shared memory between the CPU, GPU, and accelerators.
It seems like where the industry is headed!
Windows already shows something called “Shared GPU Memory” in Task Manager. On conventional PCs, this is system RAM that Windows can make available to the GPU, typically up to about half your total RAM depending on the hardware, OEM, BIOS, and drivers.
But this is a ceiling, not a fixed reservation, and Windows can use shared system memory alongside dedicated GPU memory depending on the workload.
That’s different from what Microsoft is testing here.
The new strings specifically talk about reserved memory, and Microsoft even notes that this memory would no longer be available to other applications. In other words, Windows appears to be preparing an actual memory carve-out for graphics and AI workloads.
Also, unlike macOS, Windows is adept at different kinds of workloads while also being the best place to game on, and as such, an option to control how much of this memory you want to allocate to graphics or AI compute is a must-have.
This new feature to adjust reserved unified RAM in Windows 11 is currently under the internal name IntelligentCarveout, and if you are on the Experimental build, you can maybe enable it with ViVeTool using the feature ID 61121285.
The build also ships a new file, SettingsHandlers_UnifiedMemory.dll, which is tied to this feature.
We don’t know what the final Settings page for adjusting this setting looks like. It could be a slider, fixed presets, or workload profiles, but evidence in the build points to the option being located under Settings > System > Advanced.
I recently watched a YouTuber praise a MacBook’s 128GB of unified memory while mocking a Windows gaming laptop’s RTX 5090 for having “only” 24GB of VRAM. It’s an unfair comparison. The RTX 5090 Laptop GPU is built for gaming, with dedicated high-bandwidth memory for that job. Nobody buys it to load 120B parameter AI models, but I digress because it’s not like it can’t handle AI workloads. The Mac is just better for loading huge local models because of its much larger unified memory pool.
And that’s why RTX Spark is going to be a game changer for Windows PCs. Up to 128GB of unified memory for local AI, creative work, and gaming, all on the same chip, and still with NVIDIA’s full RTX stack including DLSS, Reflex, and G-SYNC.
A memory carve-out setting is exactly what this kind of chip needs. Running a large local model can benefit from more memory being reserved for AI and graphics workloads, while other workloads may be better off leaving more of that memory available to Windows and normal applications.
Apple doesn’t give MacBook owners a direct way to adjust how much unified memory is reserved for graphics or AI workloads.
While all present-day and upcoming chips with unified memory could benefit from something like this, it’s NVIDIA RTX Spark that made all the headlines.
Windows is already changing its scheduler, memory management, and drivers for RTX Spark
Microsoft is testing new scheduling, memory management, Prism tuning, and power management specifically for RTX Spark, calling it Workload Profile Scheduling.
RTX Spark packs a 20-core Grace Arm CPU, a Blackwell GPU with up to 6,144 CUDA cores, AI acceleration, and unified memory onto one chip. Regular Windows scheduling wasn’t built for such a chip on a portable laptop.
Also, NVIDIA CEO Jensen Huang said RTX Spark runs every Windows app NVIDIA and Microsoft have ever built, through native Arm64 support and Prism emulation for x86 software. We’ll believe the full claim once we can test it ourselves.
We also found that NVIDIA has already shipped a developer-preview GeForce driver, version 616.00, for Windows 11 on Arm, and it confirms two N1X configurations, one with 6,144 CUDA cores and one with 5,120.
RTX Spark hardware is getting closer, and Windows Latest already showed some of it
NVIDIA says RTX Spark laptops and mini PCs ship this fall from Microsoft, Lenovo, ASUS, Dell, HP, and MSI, with Acer and GIGABYTE to follow.
We are looking forward to seeing the Surface Laptop Ultra, Microsoft’s RTX Spark answer to the MacBook Pro. We also published the first images of Lenovo’s Yoga Pro 9n, and later exclusive photos of the Yoga 9n 2-in-1, the first RTX Spark convertible.
Microsoft’s Surface RTX Spark Dev Box packs 128GB into a compact desktop for developers. Interestingly, a developer already got an NVIDIA RTX 4060 running on Windows 11 on Arm, well before RTX Spark ships, using NVIDIA’s own Arm64 driver work. The setup was rough and nowhere near consumer-ready, but it proves NVIDIA’s Arm drivers work wonders.
Leaked Geekbench 7 scores for the RTX Spark N1X chip already beat AMD’s Ryzen AI Max+ 395 and Intel’s Core Ultra X9 388H in multi-core performance, on hardware that isn’t even final yet.
Windows is already getting better this year, and once PCs with RTX Spark launch, a unified memory control in the OS would be icing on the cake. Of course, the feature doesn’t appear to be limited to RTX Spark, as it’s found in the Experimental “Future Platforms” branch, and other supported unified-memory chips could benefit from it as well.
Microsoft hasn’t spoken a word about IntelligentCarveout yet, but between the references we found in NVIDIA’s internal documentation, the new Windows feature flag, SettingsHandlers_UnifiedMemory.dll, and the explicit “Reserved memory for accelerators” strings, it’s clear that Microsoft is experimenting with giving Windows 11 more control over how unified memory is reserved for graphics and AI workloads.
The post Windows 11 will let you decide how much memory goes to graphics and AI on PCs with unified memory, and it’s a big deal for gaming appeared first on Windows Latest