A Guth Labs publication ↗
Back to News

GPU Task Manager 0.4.1 makes local-AI memory reporting less misleading

GPU Task Manager 0.4.1 makes local-AI memory reporting less misleading

Scott Hanselman’s gputaskman released version 0.4.1 on November 13, adding Windows Intel activity reporting and changes that distinguish shared system memory from dedicated VRAM. The release also identifies loaded Foundry Local models on Snapdragon, improves custom model-path detection with explicit provenance, and fixes integrated-GPU process-memory reporting. Release notes

The practical purpose is to explain what occupies a local GPU, not just display one utilization percentage. The MIT-licensed project connects process telemetry with runtime and model evidence. Its documentation explicitly distinguishes measured counters from estimated model classifications and memory breakdowns. A mapped model file alone does not prove that its weights are resident on the GPU. Project documentation

That distinction matters when an integrated GPU borrows system RAM or an inference process mixes weights, cache and other allocations. It also limits what the display can safely claim: the new NVIDIA TCC-mode path still carries a hardware-testing caveat, and the documentation says Intel Arc B580 behavior has been covered by mocked counters rather than validated on Intel hardware.

Guth analysis: Local-model builders need trustworthy observability as much as smaller quantizations. A tool that labels inferred memory attribution instead of presenting every estimate as measurement can help distinguish model pressure from unrelated desktop usage. This release is not evidence that a model runs faster or fits into less memory.

Evidence limitation: Release and documentation were directly read; no installation, hardware test or independent verification of version 0.4.1 was performed. GitHub displays November 13 at 06:16 without an explicitly exposed timezone in the retrieved text.

Sources and citations

Revision history

Published .