Axon NPU changelog
This page tracks changes and updates as compared to the latest official release. For more information refer to the following section.
For the list of Edge AI Add-on changelog information, see Release notes.
Changelog
See the list of changes for a specific release of the Axon NPU software.
Release 1.3.0 02 Jul 2026
Added:
Compiler release 1.3.0
Support depthwise and 2D convolutions for dilation (thorugh Space_to_Batch and Batch_to_Space operators).
TFLite Compiler - Generic classification labels are now added for
just_labelsandlast_layer_vectorslabels type if you do not provide it.
Fixed the following items for TFLite Compiler:
Handling of ‘NewShape’ attribute in the RESHAPE Operator.
Classification labels do not require test vectors to be populated in the model header file.
Static assert on the interlayer buffer size now uses “<=” instead of “<”.
Release 1.2.1 05 May 2026
Added support for
v3.3.0tag ofsdk-nrf(SHA-1 ba167d9f3d).Note
The Add-on was tested and built with SDK toolchain version v3.3.0. You should use this version for stable builds.
Removed RRAM const-lat work-around (compatible with Bluetooth LE).
Release 1.2.0 16 Apr 2026
Added:
Compiler release 1.2.0
Support for
v3.3.0-preview2tag ofsdk-nrf(SHA-1 ede152ec21).Note
The compiler was tested and built with SDK toolchain version v3.2.0. You should use this version for stable builds.
Support for multiple outputs in a model.
Support CPU operator
RESIZE_NEAREST_NEIGHBOR.static_assertin compiled model header files to verify that the interlayer buffer is allocated enough space to accommodate the model.Compatibility check so that models report a minimum supported Axon version, preventing models compiled with new features from being run on an older version of the driver that does not support these features.
Option to print a histogram of bit differences between Axon inference and TFLite inference.
Fixed:
Quantization multiplier misapplied under some circumstances when one input to an Add operation is packed and the other is unpacked. You must recompile models to apply this fix.
Compatibility
Older model versions are compatible with the latest driver. You should recompile models using the Add operator.
New model versions are compatible with older driver versions if they do not include multiple outputs or the
RESIZE_NEAREST_NEIGHBORoperation.Older driver versions will ignore all but the first output if a model has multiple outputs.
Older driver versions will not compile models with the
RESIZE_NEAREST_NEIGHBORoperation.
Release 1.1.0 19 Mar 2026
Added:
Compiler release 1.1.0.
Support for the
v3.3.0-preview2tag ofsdk-nrf(SHA-1 ede152ec21).Note
The compiler was tested and built with SDK toolchain version v3.2.0. You should use this version for stable builds.
TFLite v2.19 as the officially supported version of TFLite. Version 2.15 should still work.
Build support for
nRF54lm20b, replacingnRF54lm20a. The new board name isnrf54lm20dk/nrf54lm20b/cpuappin the build command.
Fixed:
Fully-connected layers with up to 2048 input length and 1024 output length now work correctly with TFLite 2.19. The previous version supported up to 2046 and 512, respectively.
Sigmoid and Tanh after fully-connected layers now work correctly with TFLite 2.19.
Release 1.0.1 05 Mar 2026
Added:
Compiler release 1.0.1
Note
The compiler was tested and built with SDK toolchain version v3.2.0. You should use this version for stable builds.
Fixed:
Compiler now correctly handles a fully-connected layer followed by softmax when the fully-connected layer uses per-channel weight quantization (TensorFlow v2.19).
Known issues
There are no reported known issues for this release.
Release 1.0.0 03 Mar 2026
This is the first public release of the Axon NPU software.
Added:
Compiler release v1.0.0.
Support for v3.3.0-preview tag of the
sdk-nrf(SHA-1 ede152ec21).
Note
The compiler was tested and built with SDK toolchain version v3.2.0. You should use this version for stable builds.
Arbitrary reshape operation.
Updated:
Licenses to public Nordic license
SPDX-License-Identifier: LicenseRef-Nordic-5-Clause
Fixed:
Softmax no longer specifies unpacked output. Affects a small number of models.
Known issues
There are no reported known issues for this release.
Release 0.7.0 17 Feb 2026
Added:
Compiler release v0.2.0.
Support for the
mainbranch ofsdk-nrf(tested withSHA-1 860c808db9).
Note
The compiler was tested and built with SDK toolchain version v3.2.0. You should use this version for stable builds.
Updated:
Modified the source tree to match
sdk-edge-airepository.Improved performance in compiler and simulator when performing inference.
Fixed:
Compiler now checks for dilation on depthwise convolution and returns errors.
Known issues
Current release has the following known issues:
ADD operation with broadcast on both height and width axes is not supported when both inputs are dynamic (outputs from previous layers).
Per-layer testing in unit tests is not supported due to compilation errors caused by incompatible types.