How beginners can use the new column in Windows 11 Task Manager

I tinkered with the Task Manager in Win11 over the weekend and noticed that after the update, there are a bunch of new columns related to NPU and memory. After reading through the docs for ages and running several local models, I finally figured out how to use this stuff properly. Writing a tutorial for friends who can't make sense of those new columns.

First, let's clarify a concept. What we usually call memory (RAM) is the computer's short-term memory; every program you open shoves data into it. Previously, the CPU and GPU (graphics card) used their own separate memory areas, with the GPU's area specifically called VRAM. Later, Apple Silicon popularized the "unified memory" approach, where the CPU, GPU, and NPU share the same memory pool—flexible, but prone to conflicts. So what is an NPU? It's a chip dedicated to running AI tasks, like real-time translation or image generation—a specialized processor you can think of as an "AI accelerator."

Microsoft's move this time is adding a more refined memory scheduling and visualization system within Windows 11, allowing you to see clearly which programs are eating up memory, especially how much AI tasks are consuming. This is just like when I look at VRAM usage in medical imaging: you need to see it clearly before you can manage it.

Step 1: Check if your PC has an NPU

The requirements for Windows 11 AI+ PCs state that NPU compute power must reach at least 40 trillion operations per second, but this standard is hard for beginners to check. The simplest method: Open Task Manager (press Ctrl + Shift + Esc), click the "Performance" tab on the left, and see if there is an "NPU" item in the list.

If yes, congratulations, you can follow the full process. If no, don't panic—you can still view GPU memory usage. Some columns will be missing later, but it won't affect the big picture.

Step 2: Enable AI-related columns

The default columns shown in Task Manager actually don't include AI workloads. You need to go to the "Processes" page, right-click on the column headers at the top of the list (like the row with "Name," "Status," etc.), and a long checkbox menu will pop up.

Look for these items in the menu:

• NPU: Shows whether the process is using the NPU and how many engines are occupied
• NPU Engine: Its sub-item, providing a more granular breakdown of engine usage
• NPU Dedicated Memory: In GB, the NPU memory exclusively occupied by this process
• NPU Shared Memory: The NPU memory this process can share with other processes

Check them all, then click on a blank area of the list to close the menu. Now look at the "Processes" page again, scroll horizontally, and you'll find a group of "NPU Dedicated Memory" and "NPU Shared Memory" columns added, with separators and faint colors at the top.

Step 3: Run an AI task to verify

How do you know if these columns are actually working? The most straightforward way: open any local AI application, such as a chat software with a local model or an image generation tool, and then watch the NPU and memory columns for that process in Task Manager.

In my tests, once the AI app starts working, the NPU column jumps from "0%" to a non-zero value, and NPU Dedicated Memory also rises noticeably. A few seconds after the task finishes, the values drop back down. This dynamic change is more intuitive than any documentation.

Pitfalls Section

Pitfall 1: Can't find the new columns in the right-click menu. I encountered this too initially, only to discover that Task Manager hadn't updated. You need to check "Settings → Windows Update" and update both the system and Task Manager to the latest version. Older versions only have an overall NPU activity graph, without process-level breakdowns.

Pitfall 2: Confusing NPU and GPU columns. Some AI apps actually run on the GPU rather than the NPU, especially older model frameworks or when the graphics card performance is very strong. At this point, if you see the "GPU" column jumping while the "NPU" column stays at 0, don't panic—the software chose more suitable hardware on its own. Remember one thing: You are watching "who is doing the work," not "who I expect to do the work." This is similar to running models in medical imaging; sometimes inference on the CPU is faster than on the GPU because data transfer overhead is lower.

Pitfall 3: On computers with 8GB RAM, it always looks like memory is about to explode. Microsoft mentioned in the materials that they will focus on improving the experience for 8GB devices on Win11 in the future, and 8GB RAM is not considered a qualified configuration for AI PCs. My personal view: If you only have 8GB, the "Shared Memory" column might stay yellow or even red for a long time, because the system itself, browsers, and background services are already eating up almost all the memory. When NPU shared memory squeezes in, it easily causes lag. For such machines, I suggest putting aside local AI models for now; using cloud-based ones is a decent workaround.

Product-wise, Microsoft's new features make logical sense: First, let you see the shape of AI workloads clearly in Task Manager; next comes "refined unified memory management," adjusting reserved capacity for graphics and AI. Without the transparency of black text on white screens, users wouldn't know who stole their memory.

I am optimistic about this direction. Having worked in medical imaging for years, I deeply understand that "being visible" is the prerequisite for "being manageable." Doctors dare to use AI-assisted diagnosis only because the system lays out the decision basis on the table.

What to try next after learning this

Open Task Manager, sort the NPU Dedicated Memory and NPU Shared Memory columns for each process on your computer, and see who the memory hogs are and who is secretly running AI tasks in the background. Then close a few that you think should be closable and observe the changes in memory numbers. Once you've done this step, you'll have a brand-new understanding of "who exactly is eating your computer's memory," which is far more useful than installing various manager software. When Microsoft officially rolls out refined unified memory management features in the future, you won't feel unfamiliar with them.