Login
You're viewing the functional.cafe public feed.
  • Nov 6, 2026, 9:14 PM

    @fedops We have some 4-5 year old NPU's still churning along. While I will bemoan AI use we did use it to further optimize our inference pipeline to probably get another 2 years out of it before we need start replacing them.

    💬 1🔄 0⭐ 0

Replies

  • Nov 6, 2026, 9:34 PM

    @rridley I think the future will hold locally-run models which are quite limited to information only for a specific subject. E.g. online help for industrial HMIs. Any NPU will do, even a beefy CPU is fine.

    But the big honking rigs the universal models run on are just not cost effective when old, based on their energy consumption.

    💬 1🔄 0⭐ 0
  • Nov 6, 2026, 10:39 PM

    @fedops Yeah our use case is extremely specific so they work great - object detection for a very narrow subset of objects at the edge

    💬 0🔄 0⭐ 0