Release notes for Edge AI Add-On v3.0.0

This page tracks changes and updates as compared to the latest official release.

You can also view detailed changelog pages for:

• Axon NPU

• nRF Edge AI Lib

For the list of potential issues, see the Known issues page.

Changelog

This is the first release supported for production. It is based on the nRF Connect SDK release v3.4.1.

• Updated:

  • Axon NPU to v2.0.1. All bundled Axon models were recompiled for the updated model description structure.

  • nRF Edge AI Lib to v3.0.0. Solutions and applications built for the 2.x runtime must be re-exported with Nordic Edge AI Lab 3.0.0.

  • The nRF Edge AI Observability Library metrics, APIs, and CBOR wire format:

    • Consolidated some of the built-in metrics:

      • Merged the probability distribution and class streak distribution metrics into Class predictions distribution.

      • Replaced the prediction switching rate, probability entropy distribution, and probability top-2 margin distribution metrics with Model certainty descriptor.

    • Replaced the Kconfig options of the affected metrics with new ones.

    • Renumbered the on-wire metric IDs and changed format_version to 3.

  • The decoding script to match the updated metrics and CBOR wire format. The script now accepts only the latest wire format version.

  • The Memfault CDR transport in nRF Edge AI Observability Library, which now sends incremental observability metrics. The nrf_edgeai_obsv_memfault_collect() function now encodes and resets each observability context in a single critical section, but only if no CDR payload is currently staged.

  • The Wakeword and Keyword Spotting application observability integration to use all built-in metrics.

    • Replaced the CONFIG_MODELS_OBSERVABILITY Kconfig option with CONFIG_MODELS_OBSERVABILITY_WW and CONFIG_MODELS_OBSERVABILITY_KWS. Observability can now be enabled separately for each model.

    • Expanded the wake word model prediction into a synthetic two-class [1 - p, p] vector for the probability metrics.

  • The Data forwarder sample to use gated sensor sampling only when transport is enabled (Bluetooth LE connection is active and client is subscribed to notifications), reducing average current consumption when not connected.

• Removed the deprecated Edge Impulse data forwarder sample application. Use the Data forwarder sample instead and enable the CONFIG_DATA_FWD_PROTO_ASCII_MODE Kconfig option for Edge Impulse CLI compatibility.

• Fixed:

  • NCSDK-40932: Device Firmware Update failure on Thingy:53 in the Gesture Recognition application. The root cause is NCSDK-21790 from the nRF Connect SDK.

  • NCSDK-40250: Restored serial recovery in the Gesture Recognition application release configurations for Thingy:53.

  • System workqueue stack overflow in the Gesture Recognition application during pairing with CONFIG_BLE_MITM_AUTH Kconfig option enabled.

  • The Data forwarder host tool truncating Y-axis labels on plots of individual channels.