Neutron NPU supported operations vs. Ethos U55/65

[m3d0r]

Hi there,

I have some questions regarding the supported operations on Neutron NPU. 

As far as I can see the Neutron NPU only supports a very small subset of available Tensorflow Lite Micro operations according to the Getting Started Guide. If I convert our neural network (GELU activation, BatchNorm, MaxPooling2D, Convolution2D, GlobalAveragePooling2D, Concatenation, Softmax) for MCX54x (using the eIQ Toolkit 1.10.0 on Windows) it only places the convolutions on the Neutron NPU. If I convert the network for iMX93 (with the Ethos U65 NPU) using the eIQ Toolkit it places all operations on the NPU. I have checked the computation graphs using netron.app so I am certain that the Neutron NPU only executes the convolution operations.

Am I doing something wrong or is this intended? Are there any plans to have equal NPU operation support as the ARM Ethos U55 and U65? 

Thanks in advance!

[ZhangJennie]

Hi m3d0r 

MCX54x integrates N1-16 which is basic level NXP Neutron NPU core. it doesn't have all features that Ethos U65 has.

It supports operators like

• Conv_2D
• Depthwise_Conv_2D
• Full_Connected
• ADD
• Average_Pool_2D
• Max_pool_2d
• Reshape
• Pad
• Slice

For example, N1-16 doesn't support operators like Softmax. On MCX, it can be accelerated by power quard.

Ethos U65 is more powerful NPU which is designed for imx93 by ARM.

NXP will have next version Neutron NPU core. but so far MCX product still uses N1-16 core. 

Have a nice day,

Jun Zhang