AI news, problem solutions & tool picks
DigestX

Tool Suggestions

gputaskman Is an htop for What’s Inside Your GPU, Not Just How Full It Is

· 5 min read

A graphics card with cooling fans, lit in blue
Photo via Pexels

Task Manager tells you 15 GB of your GPU is used. It does not tell you by what. Model weights? KV cache? Your browser? The desktop compositor? That is the question gputaskman answers. The tool is the work of Scott Hanselman, a Microsoft vice president, who created the repository on November 3 under the MIT licence. Its tagline, in the README's own words, is: “An htop for what's inside your GPU, not just how full it is.” By November 8 it had 673 stars and 31 forks, and the press — Neowin first on November 6, then XDA, Tom's Hardware and Windows Report — had already picked it up.

What it shows

The device view covers VRAM, GPU activity, memory activity, PCIe traffic, temperature, power, the video-engine clock and eight history charts. The process view goes further: per-process VRAM, GPU-accessible system RAM labelled “Spill”, estimated non-resident commit labelled “Paged”, engine activity, and the detected runtime — CUDA, HIP, Vulkan, Direct3D, ggml or ONNX Runtime. The model inspector is the part that makes it different from a system monitor: it identifies which model each process holds — a language model, a decision model, an embedding model, a speech model or a diffusion model — with its architecture, quantisation, layer count, weights, estimated KV cost and buffer sizes taken from load logs. Evidence is ranked api, then mapped, then cmdline, then installed, and a model backed by two sources including api or mapped is shown as “confirmed”. Live server hints add context size, busy slots and throughput, with friendly names for Ollama and LM Studio servers. Output is scriptable with --once and --json, a diff command compares two snapshots, and multi-GPU machines get --list-gpus and --gpu N. The scope is local GPU workloads — decision models, vision, speech, LLM inference, training and fine-tuning including Unsloth, graphics and video. CPU-only, NPU-only and cloud workloads are out of scope.

The releases so far

Development has moved fast. Version 0.1.0 landed on November 3 at 04:19, 0.2.0 the same evening added Linux and WSL support, and 0.3.0 added AMD support later that night. Then three more releases followed on November 7: 0.3.1 at 01:05 added the gputaskman diff command, macOS Apple Silicon support, AMD on native Linux and a Nix flake; 0.3.2 at 06:48; and 0.3.3 at 07:24, which is the current version. The event date is November 3 — the November 6 to 8 press coverage is recirculation, not a new release.

Platforms and install

Windows 10 and 11 64-bit, NVIDIA and AMD on a current WDDM driver, are covered by portable ZIPs for x64 and ARM64 that need no Python. Native Linux covers NVIDIA through NVML and AMD through the amdgpu driver with no ROCm needed, in a portable tarball that needs glibc 2.35 or newer — Ubuntu 22.04 or Debian 12 and up — and works natively, in containers and in WSL. In WSL, per-process VRAM is matched from the Windows host through vmwp.exe, with matches marked with an asterisk and guesses with a question mark, and the owner is inferred, never measured. macOS Apple Silicon is clone-and-run only, with uv run gputaskman — there is no portable macOS download yet — and because Apple Silicon uses unified memory there is no VRAM to split and no spill or commit counters; per-process GPU memory comes from footprint, which needs sudo for other users' processes. NixOS gets a flake with nix run path:. Install with uv tool install gputaskman or pipx install gputaskman, then run gputaskman — or the shorter alias gtm — on Python 3.11 and up. On portable Windows, extract the entire ZIP keeping the _internal folder beside gputaskman.exe. The first run needs internet, but no CUDA Toolkit, Visual Studio, admin session, API key or model download is required.

The numbers check out — mostly

The README documents its estimates against real measurements, and the checks are impressively exact in the simple cases. On llama.cpp build 12061 with CUDA 13.4 on an RTX 4080 SUPER, the GGUF estimate for Qwen3-0.6B Q8_0 GPU weights was 604.15 MiB, and the runtime log said 604.15 MiB — an exact match. The KV estimate for the same model at 8,192 context in f16 was 896 MiB, and the log said 896 MiB. A live example with a llama.cpp ROCm server running Qwen3-4B-Instruct-2507 Q4_K_M showed 3.95 GB of VRAM as 2.49 GB of weights plus 1.21 GB of f16 KV plus compute buffers. But the README is honest about the misses too. The Gemma-3-270M KV estimate at 4,096 context was 72 MiB against 49.5 MiB in the log — the estimate is an upper bound, because Gemma 3 uses sliding-window attention. And in a cautionary example, raising a llama-server context from 4,096 to 32,768 estimated 504 MB more KV, but measured dedicated memory grew by only 83 MB, again because of sliding-window attention. The estimates are good; they are not measurements.

What it does not do

The README states the limits plainly: “Early, experimental software. Device and process counters are measurements; classifications and GGUF-derived breakdowns are estimates. This does not dump VRAM or enumerate a game's textures.” Estimates assume llama.cpp-style placement when no logs are available, and architecture, partial offload, KV quantisation, sliding-window attention and multiple models per process all change the results; a mapped file does not prove a model is resident on the GPU. Builds are currently unsigned, with each ZIP shipping a .zip.sha256 companion, and there is no installer. WSL process owners are inferred or guessed, never measured. Multi-GPU support is covered by synthetic tests, not yet validated on multi-GPU hardware; AMD attribution has unit tests only; LM Studio is fixture-tested only; WinML and Copilot+ are not validated; bare-metal Linux on NVIDIA is not yet validated; and Intel GPUs are not yet supported. On privacy, the tool reads local process metadata, model headers, load logs and local server responses, does not dump weights and sends no telemetry — but its output can contain paths, command-line flags, model names and local ports, so redact it before posting publicly.