Microsoft VP open-sources GPU Taskman, exposing VRAM use per process

Microsoft vice president Scott Hanselman has released gputaskman, a free open-source GPU task manager that shows exactly which process consumes how much VRAM on Windows, Linux and macOS. The tool, shared under the permissive MIT license as "experimental software," also breaks out AI and large language model workloads, including model weights and KV cache sizes.

The release version documented so far is v0.3.0, available as portable Windows downloads. That is a low version number, and the author himself labels the project experimental. Buyers of nothing, users of everything: the price is zero, but the maturity is equally early.

What does the tool actually show?

Hanselman framed the motivation plainly. "There's a lot going on in my GPU," he wrote on social media. "I'm making a GPU Taskman to get insights to where my GPU memory is going."

The name states the function. The scope is broader than a simple memory meter. According to the project's GitHub repository and the screenshots Hanselman published, gputaskman reports:

• per-process GPU usage;
• VRAM consumption per process;
• active runtimes;
• live-streaming charts of GPU load, VRAM use and temperature.

One of Hanselman's own screenshots identifies the desktop windows manager as the most VRAM-hungry application on his RTX 4080S system. That detail carries a small sting. A core operating system component, not a game or a model server, topped the memory list on a modern high-end GPU.

How does it handle AI workloads?

This is the part that distinguishes gputaskman from the task managers shipped with operating systems. A GitHub-embedded image shows the llama-server.exe process selected in the tool's list. The breakdown at the bottom of the window identifies the model in use as Qwen3-0.6B and itemizes its memory footprint:

• model weights;
• KV cache size;
• parameter count;
• further runtime details.

For developers running local models, this granularity answers a real question. Standard system monitors show only that an inference server holds memory. They do not show whether the KV cache or the weights dominate the allocation, which is the information needed to size a GPU for a given model.

What did hands-on testing find?

Testing of the portable v0.3.0 build on a laptop with an RTX 4060 produced a quiet first result. On launch, the tool showed no tasks using the discrete GPU, because none were. The test then ran the Unigine Valley benchmark to wake the RTX 4060, after which the tool tracked it.

One discrepancy emerged. The live-streaming charts of GPU load, VRAM use and temperature appeared more prominent in the test screenshot than in Hanselman's published examples. The likely explanation is mundane: Hanselman may have captured an earlier build. Version 0.3.0 of an experimental tool will not match every screenshot circulating alongside it.

Could this reach Windows itself?

Hanselman hinted at more in social media comments. Expanded GPU monitoring capabilities could come to Windows as standard, he suggested. That is a hint, not a roadmap. No date, no build channel and no feature commitment accompanied the remark. Microsoft has not confirmed anything.

Still, the signal matters. A vice president publicly building a VRAM inspector says the company's own tooling left a gap. If native Windows monitoring absorbs the function, gputaskman will have served as the prototype; if it does not, the MIT-licensed repository remains available for anyone to extend.

What are the caveats?

Three deserve attention. First, the author calls it experimental software, and version 0.3.0 supports that description. Second, the tested screenshots and the developer's screenshots do not fully agree, which suggests fast iteration or inconsistent builds. Third, the claim of expanded Windows monitoring is a comment on social media, not a product announcement.

What is not in question is the license and the access. The source code sits in the GitHub repository, the MIT license permits commercial and personal use, and the binaries run on three operating systems. For anyone debugging VRAM pressure from local LLM inference, the tool answers questions that Task Manager does not. Whether it stays a hobby project or becomes a Windows feature, Hanselman has published the code that shows where the memory went.