Key takeaways

  • See which process holds your GPU memory, instead of one total from a generic task manager.
  • It links GPU memory to loaded runtimes and model files, and splits an LLM's memory into buffers.
  • It runs on Windows, Linux and WSL, with macOS on Apple Silicon also covered.
  • The catches: it is early, experimental software, and model breakdowns are estimates.
  • Support varies by GPU vendor and platform, and several paths are not yet validated on hardware.

What it is

gputaskman is a terminal interface that the GitHub description calls a GPU task manager for Windows, Linux and WSL. It is released under the MIT licence, and the code is on GitHub at shanselman/gputaskman.

The problem it solves

Anyone running local AI models knows the puzzle. A standard task manager tells you how much GPU memory is in use, but not by what. The project's pitch is to show what is inside your GPU, not just how full it is.

Its README says it tries to answer whether the memory is model weights, KV cache or something else, such as a browser or the desktop.

What you get

  • A per-process view: memory and engine activity are tied to the processes using them.
  • Context: the tool connects that usage to loaded runtimes, model files and LLM buffer breakdowns.
  • Reach: Windows, Linux and WSL are covered, plus macOS on Apple Silicon.
  • Vendor coverage: NVIDIA works across platforms, and AMD and Intel are supported on native Linux.
  • Honest labelling: device and process counters are described as measurements, with classifications treated as estimates.

The catches

  • Maturity: the project calls itself early, experimental software, and model breakdowns derived from GGUF files are estimates.
  • Identification has limits: a PyTorch or CUDA library alone cannot identify a model or prove training.
  • Scope: CPU-only, NPU-only and cloud workloads are out of scope.
  • Vendor gaps: on its AMD Windows check, temperature, memory activity and PCIe readings were unavailable, and Intel Arc B580 behaviour is covered by mocked counters but not validated on Intel hardware.
  • Platform limits: under WSL, Linux cannot see its own processes' VRAM use directly, and on Linux other users' processes need elevated rights to be listed.
  • Distribution: the initial Windows builds have no code signing or installer, and a supported GPU with a current Windows driver is required.

What we did not test

We did not run gputaskman, check its readings against a real GPU or try it under WSL. Every claim here comes from a literal quote in the project's README or the GitHub API, and the live checks mentioned are the author's own, not ours.

Verdict

Editors' view: this is for people who run local models or GPU-heavy work and want to know what is using their memory. If you only need a single utilisation figure, your existing tools are probably enough. Treat its model breakdowns as informed estimates, and check that your GPU and platform are among those it supports.

Repository
shanselman/gputaskman
Python · MIT
View on GitHub ↗

Sources

Need help adopting this, or have a software project in mind?

Tell us what you are building. The initial consultation is free. Email [email protected] or follow @zolnaltd.

Start a conversation →