Your GPU memory is full. Task Manager says “12.4 GB” and nothing else. You’re running Ollama, a browser, VS Code with an AI plugin — and you have no idea which process is the culprit. This has been the state of GPU monitoring for years. On November 6, 2026, Scott Hanselman (VP at Microsoft and GitHub) shipped gputaskman: an open-source, htop-style terminal UI that breaks down GPU memory by process — and tells you exactly how much of that VRAM is your LLM’s weights versus its KV-cache.
The Feature That Actually Matters: LLM Model Inspector
Every GPU monitor shows you that memory is in use. gputaskman shows you why. The model inspector is the differentiator: for any process running an LLM, it surfaces the model type (LLM, embedding, speech, diffusion), architecture, quantization level, and layer count. More importantly, it breaks down VRAM into three buckets: model weights, KV-cache, and compute buffer.
That KV-cache line matters more than it looks. Most developers don’t realize that llama.cpp and similar runtimes reserve KV-cache at startup for the maximum context size — regardless of how much context you’re actually using. So you can be running a relatively small model and wonder why 3–4 GB of your VRAM seems permanently spoken for. Now you can see it, right there in the terminal.
What nvidia-smi Gets Wrong (and Task Manager Doesn’t Even Try)
The existing tooling is not good enough for what developers are doing in 2026. nvidia-smi gives you a point-in-time snapshot of total occupied VRAM — not which process holds what, and not what’s actually available to your inference runtime. Windows Task Manager shows GPU utilization percentage. That’s it. Neither tool warns you when VRAM overflows to system RAM, which causes a silent performance cliff that can drop inference speed by an order of magnitude.
gputaskman runs a live TUI — refreshing continuously — with per-process rows, GPU temperature, power draw, clock speeds, encoder/decoder activity, and PCIe traffic. The comparison to htop is apt: you get a full picture of the GPU at a glance, the same way htop replaced top for CPU monitoring.
Installing It Takes Two Minutes
On Windows, there’s a portable ZIP on the GitHub releases page. Download it, extract it (keep the _internal folder next to the exe), and run. No Python installation required, no CUDA Toolkit, no model download. Requirements are 64-bit Windows 10/11 and a current WDDM driver.
For Linux, WSL, or if you want to run from source:
git clone https://github.com/shanselman/gputaskman
cd gputaskman
uv run --locked gputaskman liveA few other modes worth knowing:
# One-time snapshot (useful for scripts)
uv run --locked gputaskman --once
# JSON output
uv run --locked gputaskman --json
# Custom refresh interval (every 2 seconds)
uv run --locked gputaskman -i 2Platform Status: Honest About What Works
Windows gets the full experience: NVIDIA and AMD GPUs via WDDM and NVML. Linux and WSL have NVIDIA support baked in; AMD on Linux is in a community pull request (amdgpu sysfs + DRM fdinfo) and not yet merged. macOS Apple Silicon support is also a community PR — basic metrics, still early.
Multi-GPU support is an open GitHub issue with clear community demand, but not yet implemented. Hanselman describes the project as “early experimental software.” Take that at face value: it works, it’s useful today, and it’s moving fast. The community is already contributing — which is usually a reliable signal for where a tool is headed.
Why a Microsoft VP Had to Build This Himself
The pointed observation making rounds on X is accurate: a VP at Microsoft built the GPU monitoring tool that Microsoft’s own OS team never shipped. The Windows Task Manager has been showing the same bare GPU percentage for years while the ecosystem around local AI models exploded. gputaskman fills a gap that became urgent in 2026 — and it’s MIT licensed, cross-platform, and already useful out of the box.
If you’re running any local AI workload and haven’t checked where your VRAM is actually going, this is the tool. Install it, run gputaskman live, and you’ll learn something new about your own machine in under two minutes. The community response has been immediate and positive — expect this one to grow.
