Skip to main content
LCE
Principal II
February 16, 2026
Solved

STM32N6xy: NPU for audio DSP?

  • February 16, 2026
  • 4 replies
  • 490 views

Heyho,

I'm not a SW / DSP guy, but I have to help selecting a new MCU.

Mostly we need gigabit ethernet, SAIs / I2S for audio input.

But for some future applications we might need some DSP power for online audio processing, some heavy filtering and FFT stuff.

I wonder if the N6 's NPU can help with that. In all the documentation / ads I mostly see graphics applications.

Best answer by Tuomas95

It is exactly made for audio processing (among other things), see the documentation here, FFT and FIR are supported.

4 replies

Associate III
February 17, 2026

Hello,

The N6 has Helium vector acceleration which can be used for the DSP processing.

LCE
LCEAuthor
Principal II
February 19, 2026

This does not seem to help with typical audio processing, using FIR filters, FFT, ... - or so it seems to me.

Made rather for 2D graphics stuff.

Tuomas95Best answer
Associate III
February 19, 2026

It is exactly made for audio processing (among other things), see the documentation here, FFT and FIR are supported.

DmitryR
Associate III
February 24, 2026

Hi @LCE,

 

Just like AI, DSP functions are linear combination of input data. Or I should better say that AI is like DSP but much larger. FIR is a single artifical neuron without activation function. FFT is one layer of perceptron. So an AI accelerator may easily be used for DSP. If you search Keras of Tensorflow you will quickly find there DSP functions that probably may be directly imported to STM32N6 project as CubeMX supports both Keras and TFlite.

 

Helium may also be used, but it is relatively week in compare to AI. AI accelerator in STM32N6 contains 4 convolutional accelerators with 18 16*16 MACs each which gives incredible for DSP 72 MACs per clock. That means for example that it could calculate a 63-tap FIR each clock. I am sure that it will be enough not only for audio but for at least some of SDR applications. 

 

But before you choose it I would recomment to create a CubeMX project, load models you need there and at least validate them. Even better will be to buy a kit and test it in hardware. 

 

With best regards,

Dmitry

LCE
LCEAuthor
Principal II
February 25, 2026

Thanks to all so far!

In CubeMX I found out that 1 pin is shared between SAI (data in) and ETH (clock) - and I need both.

So for now that rules out the N6...

Maybe I'll come back later to the N6 and work around that using I2S and some terrible PCB routing...

mƎALLEm
Technical Moderator
February 25, 2026

Hello @LCE ,

Please accept as solution if your question has been answered.

For SAI/ETH shared pin, you've already posted that question here.

Thank you

"To give better visibility on the answered topics, please click on ""Accept as Solution"" on the reply which solved your issue or answered your question."