STM32N6 project builds and deploys, but freezes on dev-kit when starting edge-impluse-run-impulse

DJE666 · January 27, 2026, 4:09pm

Dear Forum,

Project iD 851238 builds as expected resulting in a .zip with the required .bin and .hex files to deploy the complete application onto my STM32N6 dev-kit. I can flash the files to the kit and the application starts with the camera output shown on the display and the message “Inference not Running” on the screen.

Running the edge-impulse-run-impulse command on my Ubuntu laptop results in the following screen and the image on the dev-kit freezes until I reset the board.
STM32N6-yolo-pro-lockup-screenshot 27-1-26

Any ideas why this demo will load but not run? I have run more basic Edge-Impulse projects on this dev-kit successfully in recent weeks, so I know that the basic HW/SW setup is good.

TTFN,

DJE666

ei_francesco · February 3, 2026, 8:45am

Hi @DJE666 ,

can you run
edge-impulse-run-impulse --verbose ?

This will output additional messages, I will try to replicate on my setup.

thank you!

fv

DJE666 · February 3, 2026, 11:30am

Hello FV,

Thanks for the feedback. Attached is a screenshot of the STM32N6 dev-kit running an older project (CPU only) (left), but stalling on a new project (CPU + NPU) Right.

STM32N6-EI-Good-Bad-3-2-26

The --verbose output is realized as a serial stream, so I cannot capture a meaningful screenshot of the output. Can I attached a file to this support ticket?

Regards,

David

DJE666 · February 3, 2026, 1:03pm

Hello FV.

I rebuilt the Edge-Impulse project 851238 with image-size of 96x96 and the new impulse runs for a while on the STM32N6 dev-kit (which does not lock up). However no objects are detected and then the edge-impulse-run-impulse appears to crash.

Running the –verbose option adds nothing to the output, but does trash the format.

Regards,

DJE666

ei_francesco · February 3, 2026, 2:41pm

Hi @DJE666

the 320x320 is probably too big and won’t fit, I’m checking your project on my device, I will update you when I find what’s wrong.

regards,
fv

ei_francesco · February 3, 2026, 4:45pm

@DJE666
we found a bug in our ST inference code, the fix is under review, will be in production soon.
Thank you for your patience!

If you want to workd directly with our fw repo, you need to delete lines 124 and 125 here

github.com

edgeimpulse/firmware-st-stm32n6/blob/c88acb72b647bcb065c7429a22f978cf4514312b/edgeimpulse/edge-impulse-sdk/classifier/inferencing_engines/aton.h#L124


      
          while (nn_out_info[i].name != NULL) {
                  SCB_InvalidateDCache_by_Addr((float32_t *) LL_Buffer_addr_start(&nn_out_info[i]), LL_Buffer_len(&nn_out_info[i]));
                  i++;
          }
          #endif
          
          
result->timing.classification_us = ei_read_timer_us() - ctx_start_us;
          
          
size_t output_size = nn_out_len;
          
          
result->_raw_outputs[learn_block_index].matrix = new matrix_t(1, output_size);
          result->_raw_outputs[learn_block_index].blockId = block_config->block_id;
          
          
switch (graph_config->quant_type) {
              case kTfLiteFloat32: {
                  result->_raw_outputs[learn_block_index].matrix = new matrix_t(1, output_size);
                  memcpy(result->_raw_outputs[learn_block_index].matrix->buffer, (float *)nn_out, output_size * sizeof(float));
                  break;
              }
              case kTfLiteInt8: {
                  result->_raw_outputs[learn_block_index].matrix_i8 = new matrix_i8_t(1, output_size);

regards,
fv

DJE666 · February 3, 2026, 5:03pm

Hello FV,

Thanks for this update. I look forward to implementing my projects in the “fixed” environment.

TTFN,

DJE666

ei_francesco · February 4, 2026, 2:45pm

Hi @DJE666

the fix has been merged, should be live soon.
About model performance, you can see that precision is quite low.
With model testing you can see the accuracy is really low.
Probably the dataset is too small for 4 classed and picture are too similar, the background is more or less always the same.

You can also try to increase the image size (224x224 should be ok with nano instead of small) and the number of training cycles.

regards,
fv

DJE666 · February 4, 2026, 5:08pm

Hello FV,

Thanks for this update. Yes, I need to massively improve my data set.

Looking forward to a fresh build and successful deployment this week.

Regards,

DJE666

DJE666 · February 5, 2026, 3:27pm

Hello FV,

I’ve recompiled the project and it now runs on the STM32N6 Dev-Kit without crashing. However, despite the model showing good (if not great) Pokémon detection when running a live-classification on my laptop, the application and model now running on the STM32N6 dev-kit show zero signs of object detection.

Confidence on the laptop based live-classification for each Pokémon is around 50%, but on the STM32N6 dev-kit I see nothing at all.

Any ideas why the model is running so poorly on the dev-kit? When I first built this project for the STM32N6 dev-kit back in October 2025 (using even less training data), I always achieved Pokémon detection success on the dev-kit, even if the confidence was ~50%.

I cannot understand why detection on the STM32N6 dev-kit has ceased…

Regards,

DJE666

DJE666 · February 5, 2026, 7:52pm

Hello FV,

Still tinkering with this project, but deployment of new projects on the STM32N6 seems to have hit a dead end. It runs, but produces no results.

I am seeing a live image from the camera on the dev-kit display with either the Inference not Running or Objects 0 overlay in place, depending on whether the system is running or not.

Adding –debug does not result in a streaming output to my web browser.

I’m getting a 99% precision score during training and the Live classification of this project using the laptop and web-cam gives excellent results, but I’m getting nothing back from the STM32N6.

My old October 2025 STM32N6 project output files still run, but I simply can’t get new projects to run on the exact same hardware.

Any fresh ideas will be greatly received.

Regards,

DJE666

DJE666 · February 11, 2026, 4:02pm

Hello FV,

I think we can close this support enquiry as I have successfully built a new project (901657) running on the STM32N6 dev-kit that is successfully discriminating between Bird and Cat classes.

This projects has 100 times the training data than I was able to get for the Pokémon detection project, so perhaps that’s why I’m getting much better reults.

Regards,

DJE666

ei_francesco · February 12, 2026, 12:25pm

Hi @DJE666

that’ great, and sorry if I didn’t reply to your previous message.

Let us know if you need further assistance.

regards,
fv