Microsoft VP open-sources gputaskman, cross-platform GPU task manager with AI/LLM VRAM breakdown

Microsoft VP Scott Hanselman has published gputaskman, an open-source GPU task manager licensed under MIT, designed to provide granular visibility into GPU resource allocation. The tool runs on Windows, Linux, and macOS and displays per-process GPU utilization, VRAM consumption, active runtimes, and live-streaming charts for metrics including temperature and load. A distinctive feature is its detailed breakdown of AI and LLM workloads: when monitoring processes like llama-server.exe, gputaskman exposes model-specific information such as the Qwen3-0.6B weights, KV cache allocation, and parameter counts. The current release is version 0.3.0, with Hanselman actively soliciting community contributions via GitHub.

The tool addresses a persistent gap in GPU observability, particularly as VRAM contention grows more acute with local AI inference and large model deployment. Hanselman's public hints suggest that expanded GPU monitoring may eventually be integrated into Windows itself, though gputaskman remains independently developed open-source software. The terminal-like interface and portable distribution model position it as a lightweight alternative to vendor-specific monitoring suites.

gputaskman architecture and GPU telemetry capabilities

• Cross-platform runtime support: gputaskman operates on Windows, Linux, and macOS, abstracting GPU telemetry APIs across disparate vendor driver stacks rather than relying on platform-specific tools like nvidia-smi or Windows Task Manager.
• Per-process VRAM accounting: The tool attributes GPU memory consumption to individual processes, enabling identification of VRAM-heavy applications such as desktop compositors, benchmarks, or AI inference servers.
• AI/LLM workload decomposition: For inference processes like llama-server.exe, gputaskman breaks down memory usage into model weights, KV cache, and parameter counts, with demonstrated support for models including Qwen3-0.6B.
• Live telemetry streaming: The interface renders real-time charts for GPU load, VRAM utilization, temperature, and related metrics, with visual prominence increasing in the v0.3.0 release compared to earlier screenshots.
• MIT-licensed experimental software: Hanselman explicitly labels the project as experimental, distributing source code and portable Windows binaries while inviting pull requests for community-driven expansion.

Why it matters

For Linux users and developers running local AI inference with tools like llama.cpp or Ollama, gputaskman fills a tooling gap by exposing fine-grained VRAM attribution that existing system monitors typically lack. The cross-platform MIT license ensures the tool can be packaged by distributions and integrated into developer workflows without proprietary restrictions, while its open development model may accelerate support for additional GPU vendors and inference runtimes beyond NVIDIA's ecosystem.