For computers without a discrete graphics card, daily graphics processing, AI computations, and system tasks must share a single block of memory. Currently, how this memory is allocated is entirely determined by the Windows system itself. However, as tech giants like Apple, Qualcomm, and NVIDIA begin building AI PCs around unified memory pools (which could reach hundreds of GB in the future), the method of memory allocation is becoming increasingly important.
A recent Windows 11 preview build contains clues suggesting that Microsoft is preparing to give users control over how system memory and VRAM are divided within the unified memory pool. With this feature, players will be better equipped to handle both gaming and AI tasks simultaneously, especially on upcoming unified memory devices like NVIDIA's RTX Spark platform.
The Windows 11 version 29648.1000 released on August 17 introduced a new feature called "Smart Allocation," although Microsoft's update logs made no mention of it. Well-known Windows leaker @phantomofearth confirmed to Windows Latest that the code in this build indeed mentions reserving dedicated memory for accelerators, graphics processing, and AI.
This means users can manually allocate a portion of the unified memory exclusively for games and resource-intensive AI applications. This reserved memory block cannot be accessed by any other programs.
Similar statements have actually appeared in NVIDIA's official documentation before. If this feature is eventually released to the public, it will completely revolutionize how people optimize performance on unified memory chips.
Unified memory has traditionally been used in Macs, some laptops, PC handhelds, and other systems that rely on integrated graphics rather than discrete GPUs.
Depending on the task, performance can vary significantly compared to discrete graphics cards with dedicated VRAM. Although heavy graphical tasks can borrow system memory when VRAM runs out, this approach often leads to a substantial drop in performance.
Running local AI models requires massive amounts of high-speed memory, turning high-end graphics cards originally designed for gaming into hot commodities for AI workloads. Meanwhile, computers built specifically for local AI models are featuring increasingly larger unified memory capacities. With the boom in local AI, even Macs equipped with 24GB or more of unified memory often face memory constraints. The solution offered by Microsoft and NVIDIA is the RTX Spark device, which provides up to 128GB of ultra-large memory.
NVIDIA's dominance in the GPU sector isn't just due to its hardware chips; its powerful software ecosystem plays a crucial role as well. CUDA programming, DLSS super-resolution, Reflex low-latency technology, and the entire RTX suite have become core selling points that are even more attractive than raw hardware specs. Now, NVIDIA has ported this software ecosystem intact to the integrated graphics space, giving RTX Spark an inherent advantage over AMD, Qualcomm, and Intel, as these three competitors currently lag far behind NVIDIA in terms of software ecosystems for AI and gaming performance.
When NVIDIA first announced the RTX Spark chip, it boldly claimed that its supported addressable memory capacity exceeded that of any of its own GPUs. Thanks to a dedicated allocation feature, users can directly assign a large portion of this massive 128GB memory pool to graphics and AI tasks. As a result, the GPU in the RTX Spark can ultimately access far more memory than the 32GB of dedicated VRAM found in the top-tier RTX 5090 graphics card.
Of course, this shared memory pool based on LPDDR5x cannot match the speed of the GDDR7 memory in the RTX 5090, so its advantage lies in capacity rather than speed.