Microsoft Vice President and Member of Technical Staff Scott Hanselman has released an open-source utility called GPU Taskman to provide detailed visibility into graphics processing unit memory consumption. The terminal-based tool, known as gputaskman, offers more granular data than the standard Windows Task Manager by combining VRAM usage, GPU-accessible system memory, individual process activity, and hardware telemetry into a single interface.
The utility displays real-time statistics including GPU utilization, temperature, power consumption, clock speeds, memory activity, and PCIe statistics. It is designed to distinguish between various workloads such as desktop compositing, web browsers, games, and AI applications. This functionality allows users to troubleshoot excessive VRAM consumption or investigate why a GPU might be relying heavily on system memory.
A primary feature of the tool is its ability to identify local AI workloads and provide specific details regarding model architecture. According to the project's GitHub documentation, GPU Taskman can identify supported AI models and report on their quantization, parameter counts, model weights, and estimated KV-cache requirements. The utility currently recognizes several runtimes and applications, including llama.cpp, Ollama, and LM Studio.
To provide this level of detail, the software combines operating-system GPU counters with process metadata, model files, runtime logs, and local APIs. The developers noted that some model-specific values are estimates rather than directly measured GPU allocations. Hanselman cautioned that GPU-accessible system RAM usage does not necessarily indicate VRAM overflow, and the utility cannot identify individual game textures or determine precisely which resources have been evicted.
GPU Taskman supports NVIDIA and AMD GPUs on Windows 10/11, as well as Linux, WSL, and Apple Silicon macOS systems. The latest v0.3.5 release, published on November 10, added Intel GPU support on Linux, CPU and system RAM monitoring, and options for recording GPU statistics. The software is distributed under the MIT license and is available via GitHub or PyPI.
"I wanted better visibility into what was consuming my GPU memory," Hanselman wrote on X regarding the motivation for the project.
