Lacksan community

Lacksan HP Performance Forum

Workstation tuning, measured and written down.

Read-only preview

GPU / Local AI

VRAM is the binding constraint for local models, not compute

Pinned Started · 1 replies · 2411 views

  1. Documented fact

    For anyone arriving from the general PC side: on mobile workstations the thing that decides whether a local model runs acceptably is almost always whether its working set fits in VRAM, not how fast the GPU is.

    Once you spill out of VRAM the transfer over the bus dominates and the generation rate collapses. A slower GPU with more memory will frequently beat a faster one with less on the same model.

    • Check the actual VRAM figure for your SKU, not the family. Mobile parts vary within the same model name.
    • Account for the display using VRAM as well, particularly with external monitors attached.
    • Quantisation changes the working set substantially and is usually the first lever worth pulling.
  2. Hypothesis

    Related and untested: on a docked mobile workstation driving two external displays, I would expect the display allocation to push a model that previously fit just over the edge, producing an apparently random slowdown that correlates with being at the desk rather than with anything about the model.

    Plausible mechanism, no measurement. If anyone has a docked ZBook and a model sized near the boundary, this is a straightforward thing to check.

Replying is not available in this build. Accounts, posting and moderation arrive in the next phase; until then, corrections go to admin@lacksan.com.