you are viewing a single comment's thread
view the rest of the comments
[–] 7 points 6 months ago (3 children)

The newer CPU generations come with cores optimized for this stuff (referred to as an NPU). It actually seems to work fairly well for the kind of model you'd run locally.

Barring that, a typical laptop dGPU will also work, although not super efficiently since they often don't exceed 8 GB of VRAM and thus can't run most models without partially offloading them to the CPU.

Of course a laptop with a dGPU and NPU cores will make the offloading less painful. So yeah, workable for most reasonably-sized models.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 4 points 6 months ago (2 children)

    Models can split loads across a discrete GPU and CPU/RAM.

    Its not as fast as if you can load it all in the GPU, but it gives you more options. Its been quite common for a long time.

  • source
  • parent
  • hideshow 2 child comments