DevHeads IoT Integration Server•12mo ago

Is an SD Card Alone Sufficient for System Memory in an ESP32-S3 Design?

Hello @here I'm designing an electronic board around the ESP32-S3.I want to use only the SD card for the system memory. The SD card alone is enough or should I add another type of memory?

FatogomaAPP•1/8/25, 10:59 PM

Also I want to use camera through DVP interface (ESP32-S3 support camera). Which pins can be use to connect the camera?

daleonpzAPP•1/8/25, 11:06 PM

Well I guess that depends on the application..

daleonpzAPP•1/8/25, 11:06 PM

yes it 's for a computer vision application.

daleonpzAPP•1/8/25, 11:06 PM

Can you explain more about your intended use? Is it just a personal project? What model are you running? Etc.

@wafa_ath and @undefined.elf have worked on similar systems and should be able to walk you through the process

daleonpzAPP•1/8/25, 11:06 PM

The project is about to detect whether a person is sleeping or not and many others behaviors. Computer vision models with openCV,tensorflow,mediapipe and other libraries will be used.My job is to make a board that will be able to run all the models.Also connectivity is required.That's why I chose ESP32-S3.

daleonpzAPP•1/8/25, 11:06 PM

if you want to use CV with Deeplearning or nn, you should take flash, ram and inference time into account.. i could run yolov5 with 4 classes in a esp32-s2 using 96x96x3 images, but it was quite a challenge because of the RAM and processing constrains. My inference time was 350-500ms. but if you want to use bigger images, and recognize objects at differnet scales, that could be quite challenging

daleonpzAPP•1/8/25, 11:06 PM

In my personal experience using edge impulse is the best way to train and deploy Ml model on esp32 , ( i assume that u are using camera for the detection)

daleonpzAPP•1/8/25, 11:06 PM

Ow i see that u mention that, i didn't notice it

daleonpzAPP•1/8/25, 11:06 PM

Ok dear. By choosing an external flash memory and PSRAM with a good performance, this can improve the performance?

daleonpzAPP•1/8/25, 11:06 PM

yes I'm using a camera.Ok I will suggest edge impulse to the model developer. I have done a prototype of the project on a raspberry pi.But now we want to make a customized board that satisfy our requirement.

daleonpzAPP•1/8/25, 11:06 PM

daleonpzAPP•1/8/25, 11:06 PM

chek this https://github.com/LUXROBO/edgeimpulse-esp32-firmware

GitHub

GitHub - LUXROBO/edgeimpulse-esp32-firmware: Edge Impulse firmware ...

Edge Impulse firmware for the Espressif ESP-EYE(ESP32) Development board - LUXROBO/edgeimpulse-esp32-firmware

daleonpzAPP•1/8/25, 11:06 PM

Thanks @wafa_ath

daleonpzAPP•1/8/25, 11:06 PM

By choosing the right memory size can this improve the performance of my system? Or I have to consider choosing another MCU?

daleonpzAPP•1/8/25, 11:06 PM

@Fatogoma here's a blog @daleonpz just published, ironically: https://baremetallics.substack.com/p/running-yolov5-on-a-esp32-microcontroller

Running YOLOv5 on a Microcontroller

Optimizing YOLOv5 for ESP32: A Guide to Running AI on Microcontrollers

daleonpzAPP•1/8/25, 11:06 PM

I mean there are already some modules like this

daleonpzAPP•1/8/25, 11:06 PM

it can be found with very good memory like 8mb sram and 16mb flash

daleonpzAPP•1/8/25, 11:06 PM

also it is easier to implement into your project

daleonpzAPP•1/8/25, 11:06 PM

Ok. I'll do some research about this module.Thanks.

daleonpzAPP•1/8/25, 11:06 PM

when possible ,choose the mcu that already fulfils your requirements, because you avoid the overhead of managing the connection, but as always is cost of HW vs cost of SW development, the more restrictive the hw is the most expensive and risky the SW will be. sadly I cannot give you a concrete answer