Testing the ability to customize GPU allocation with Windows Unified Memory to meet memory demands for AI processing

The other day, we introduced the topic that Microsoft is testing a feature that allows users to customize the amount of unified memory allocated to the GPU in Windows, and this is in response to the rapidly increasing demand for memory for AI processing.

Judging by the series of posts by X user Xeno mentioned in the article, this is a setting item that was confirmed on "Build 29648.1000" distributed on the Windows Insider Program's Experimental Channel (Future Platforms), and considering the numbers after the setting change, it can be determined that it is a setting item that falls under "Dedicated GPU Memory", which prioritizes allocation to the GPU.

In Windows, memory is classified into three categories: "CPU Reserved Memory," "Shared GPU Memory," and "Dedicated GPU Memory." Dedicated GPU Memory is memory that can be occupied by the GPU, and this applies to memory installed in GPU cards, etc., or for example, in the case of a structure that is integrated with the CPU like iGPU, a certain amount of main memory is allocated as Dedicated GPU Memory and used for screen drawing etc.

After excluding the reserved memory in this way, approximately half of it is allocated to CPU Reserved Memory and Shared Memory, respectively.

The former is memory dedicated to the CPU, and the latter Shared GPU Memory is shared by the CPU and GPU, and the allocated amount can be changed depending on the purpose. The reason why Shared GPU Memory is about half of the main memory is to ensure stable operation of the CPU and is a basic specification of Windows...as explained in the previous article.

GPU allocation within Shared GPU Memory fluctuates in usage, but some processors have a function that increases the allocation of Dedicated GPU Memory to ensure stable GPU operation.

For example, in the case of an APU like ``AMD Ryzen AI Max+'', a function called ``VGM (Variable Graphics Memory)'' can be used using the Unified Memory area. This is a function that customizes Dedicated GPU Memory mainly for AI use. In the example in "FAQs: AMD Variable Graphics Memory, VRAM, AI Model Sizes, Quantization, MCP and More!", in a system with 128GB of Unified Memory, the Dedicated GPU was originally 512MB. After customizing the memory allocation, 96GB will be allocated to it, and the remaining 32GB will be allocated to the CPU and Shared, each with 16GB each.

What we can see here is that when Unified Memory is allocated to Dedicated GPU Memory, the remaining memory is treated as main memory and is divided into ``50% CPU and 50% Shared'', which is the basic behavior of Windows mentioned above. Basically, it seems that the principle remains the same even with functions like VGM.