Is Cube-AI based on ARM CMSIS-NN?

[RKopp.12]

How Cube-AI is different from ARM CMSIS-NN and ARM NN SDK?

[jean-michel.d]

​Hello,

Currently X-Cube-AI 3.3.0 provides only a support for the floating point model (float32). It provides its own STM32 optimized libraries (mainly based on a ARM Cortex-M with FPU single precision). ARM CMSIS-NN package is only for the fixed-point support (q8, q16) and to use the Cortex-M DSP/SIMD extensions. Future X-Cube-AI release will have a fixed-point support which will be based on the low-level ARM CMSIS-NN functions.

X-Cube-AI vs ARM-NN-SDK?

ARM-NN is actually only an inference engine for CPUs, GPUs and NPUs (ARM Cortex CPUs, ARM Mali GPU,...), it is currently not compatible with Cortex-M microcontrollers. There is not an official Cortex-M support through the ARM-NN-SDK (or derived) on the top of the CMSIS-NN library (model translator or convertor point of view) to generate an optimized C-model implementation.

Best Regards,

Jean-Michel  

[RKopp.12]

Thanks Jean. Cheers!

[DUlia]

What about Cube-AI vs CMSIS-NN from performance perspective? CMSIS NN claims 4x runtime improvement due to, as you've already said, fixed point calculations. On the other hand, their library support seems messy: you need manually add layer by layer in your source code. I'd like to see a simple comparison of memory and FPS for MNIST model with CubeAI and CMSIS NN backends.

Thank you.

[RKopp.12]

Hi ST, Recently Google released experimental version of TensorFlow Lite for ARM micro controllers. Can we have ST AI framework integrated with CMSIS-NN as a backend and TF-Lite as a front end please? This will reduce the fragmentation of AI frameworks on micro controllers. Get Outlook for Android<>

[JNord]

In X-CUBE-AI 4.0 there is support for quantized TensorFlow Lite / Keras. But will it use the actually use the SIMD instructions of an M4F if using 8-bit quantization (like CMSIS-NN)?