A game equipped with a function that guides gamers through deep learning.
Functions such as visualizing the external voice data processing process
Winner of STM32MP1 Complete Conquest Quest Challenge


The STM32MP1 Quest Challenge, open to anyone interested in electrical and electronic system design, ended on June 22.

The 'STM32MP1 Complete Conquest Quest Challenge', hosted by STMicroelectronics Korea and organized by e4ds Make, was held from April 13 to June 14.

This competition, which involved developing an original application using the 'STM32MP157C-DK2 Discovery Kit' equipped with the STM32MP1 board, consisted of a total of seven quests, one per week. The final quest to complete the application was conducted over a special two-week period.

You can only challenge the next quest if you pass the previous quest, and participants who completed all the way to the final quest received prizes such as the Galaxy S20, Galaxy Watch, and Galaxy Buds according to the ranking selected by ST Korea.

A total of 238 engineers took on the STM32MP1 Complete Conquest Quest Challenge, and only 14 completed all the quests. On June 22nd, the final rankings were announced and the four engineers who will receive the Galaxy Watch were confirmed.

e4ds News asked engineers 'Chiphead' and 'Mason' about their feelings on winning the competition, the challenges they faced, and the experience they gained as engineers.

Developed by Chiphead engineers, the ' Deep Learning-based AI Bird Game ' is a game in which you control an 'AI Bird' that soars upward whenever you touch the screen and try to pass through narrow spaces as much as possible. When you turn on the AI function, the learned deep learning tells the gamer where the AI Bird should be.

To implement this AI feature, which becomes more accurate the more a gamer plays, Chiphead engineers lightweighted the deep learning model and focused on accelerating the deep learning speed as much as possible by utilizing 3D GPU and NEON SIMD.



Q1. Congratulations on winning the STM23MP1 Complete Conquest Quest Challenge. Could you briefly introduce yourself?
A. Hello. I am a 40-year-old engineer and Chiphead who researches and develops embedded system software for mobile devices.


Q2. How did you find out about this competition? Have you always been interested in electronic circuit design?
A. I found out through e4ds email subscription.


Q3. Please explain the task you developed.
A. I wanted to implement training and inference apps that can run on embedded boards using AI network models. I used ARM NEON SIMD to accelerate training speed.


Q4. Are there any particular difficulties that come to mind during the development process?
A. Since there were not many deep learning training examples on the development board, I had no choice but to repeat many development mistakes while developing related projects.


Q5. How do you feel about using the STM23MP1 board and Qt? Was it difficult or challenging?
A. I have used Qt and found that the development toolkit has many features and is well-documented, which has helped me increase my productivity in developing GUI apps.


Q6. If a competition like this Quest Challenge were held again, would you participate?
A. I do not plan on participating in any competitions that do not utilize development boards with built-in AI-related NPUs.


Q7. Is there any part of the competition that you would like to see improved?
A. I think the evaluation would have been more fair if all applicants had developed under the same hardware conditions without using additional hardware.

' Voice signal frequency processing and filtering then sending to QT app ' developed by Mason Engineers is a function that visually displays on the display of the STM32MP157C-DK2 development board how external voice data transmitted to the A7 core of the STM32MP1 through the microphone module is processed.



Q1. Congratulations on winning the STM23MP1 Complete Conquest Quest Challenge. Could you briefly introduce yourself?
A. Hello. I am a college student in my 20s who is new to embedded systems. I recently took a leave of absence from college in the US and returned to Korea to live there.


Q2. How did you find out about this competition? Have you always been interested in electronic circuit design?
A. In fact, even 6 months ago, I didn't know anything about embedded systems. I only had a basic knowledge of electronic circuits. During my break from school, I thought I should learn something, so I started studying AI on my own. As a result, I naturally became interested in electronic circuits and came into contact with MCUs.

At first, I started with Arduino, and now I am focusing on studying STM32 MCU. While searching about STM32 MCU, I came across e4ds by chance, and also found out about this competition. In a way, I started studying by chance, and found out about the competition by chance.


Q3. Please explain the task you developed.
A. When I first saw STM32MP1, I thought a lot about “How can I use this properly to say that I used it properly?” The A7 core has the characteristics of high performance and high power, but the M4 core has the strength of moderate performance and low power.

“Can’t we greatly improve battery efficiency with the M4 core?” I investigated and found that STM32MP1 supports a function called ‘C Sleep’. I thought that if the M4 core could wake up the A7 core in C Sleep state only when necessary, it would be possible to improve battery efficiency.

So, we implemented voice recognition on the M4 core using Cube AI provided by ST, and configured it to wake up the A7 core on the M4 side when a specific word is heard. △My goal was to extract data from the microphone using ADC within the project, △apply sound pre-processing, △infer words using a speech recognition model, △wake up the A7 core by executing C Sleep, and △show the sound spectrum in a Qt app.

Unfortunately, due to limitations in knowledge and time, I was unable to complete the sound processing and output to the Qt app using RPMsg. I will try again when I get the separately purchased microphone module.


Q4. Are there any particular difficulties that come to mind during the development process?
A. The STM32MP157C-DK2 board I received while working on the project had a total of 256KB of memory that could be used by the M4 core, including SRAM1 and SRAM2. Since I had to solve both flash and RAM with 256KB, I unfortunately had to adjust the 'optimization level' along with the linker script. I thought that 256KB might be insufficient if I were to add more libraries later.

As a workaround, there was a way to map external flash to the memory map via QSPI, but unfortunately, the QSPI pin is only available on the evaluation board. Also, I didn't know about the engineering mode on the board, so I think the development process would have been easier if I had known about it. Because of this, I had to reset the entire board several times after each execution of the M4 core.

Lastly, embedded Linux and Qt were difficult for me because they were my first exposure to them in this competition, but it was still an enjoyable experience because I learned a lot from them.


Q5. How was your experience using the STM23MP1 board and Qt? Was it difficult or challenging?
A. I was new to embedded Linux and app development, so everything was new and difficult for me. In fact, I was able to implement almost everything with Qt. In particular, it was really easy to develop QML UI through QT Designer. Later, I imagined making a restaurant serving robot using STM32MP1 board and QT app.


Q6. If a competition like this Quest Challenge were held again, would you participate?
A. I learned a lot and had a lot of fun during this competition, so if this competition were to be held again, I would like to participate right away. If I participate next time, I would like to interact with more people than I did now.