Theatre

Neural Networks: From GPU to MCU

Watch Now

37:37

EOC 2023

Login to Watch

Louis Gobin

Louis Gobin is a Product Marketing Engineer supporting Artificial Intelligence (AI) solutions in STMicroelectronics Americas Region and has held this position since February 2022. Gobin started his career as a machine-learning applications engineer for Cartesiam, based in New York. After ST acquired Cartesiam, Gobin moved to Santa Clara, CA in 2021 to take on the role of Field Application Engineer. He transferred into Product Marketing Engineer position in early 2022.

View full profile

Markus Mayr

Markus Mayr leads the general-purpose microcontroller and microprocessor marketing team at STMicroelectronics in the Americas region. He is responsible for business development and design-in activities at key accounts and mass market customers. With extensive experience working with customers globally, Markus has contributed to ST becoming a leading MCU manufacturer in the Americas. His focus is on microprocessors, wireless products, and edge-AI solutions. Markus Mayr holds a bachelor’s degree in electrical engineering from Furtwangen University in Germany.

View full profile

Allen Ren

Allen Ren is a Field Applications Engineer supporting Artificial Intelligence (AI) solutions in STMicroelectronics Americas Region and has held this position since joining the company in December 2019. Allen Ren was born in Beijing, China and is currently based in Santa Clara, CA. He graduated from San Francisco State University with a Bachelor of Science in Electrical Engineering and a Master of Science in Embedded Electrical and Computer Systems in 2019.

View full profile

STMicroelectronics

Electronic circuits form a mostly invisible part of the world we live in today. These circuits are present in microchips or just "chips", which combine the latest advances in performance, intelligence and, efficiency. Hundreds or thousands of these chips are integrated into each of the millions and billions of electronic devices people across the globe interact with every day. It is in this unseen realm that STMicroelectronics creates the sparks that animate the products we use at any given moment. Our technology is found everywhere microelectronics makes a positive contribution to people's lives. Chips from ST embedding the most advanced innovations are an essential part of products as diverse as cars and key fobs, giant factory machines and data center power supplies, washing machines and hard disks, and smartphones and toothbrushes. We help our customers make these devices more intelligent, more energy efficient, more connected, safer and more secure.

Website LinkedIn Twitter

Session Resources

Download Slides (PDF)

Topics Covered

Hardware AI/ML Microcontroller

Theatre

Neural Networks: From GPU to MCU

Louis Gobin + 2 more

37:37 EOC 2023

Louis Gobin

37:37

Join this webinar to understand step-by-step how to port your machine learning model from your preferred development environment to an MCU. We will go over a few of the pipelines available to you and cover:

Graph optimization
Model quantization
Memory optimization

1 / 4

Please log in or create an account to test your knowledge and see the answers.

Why would a team choose STM32Cube.AI instead of NanoEdge AI Studio when moving a model from a GPU development environment to an STM32 microcontroller?

A Because STM32Cube.AI ingests models from common frameworks (TensorFlow, PyTorch, ONNX, etc.) and optimizes them for STM32 without changing the model's internal logic by default.

B Because STM32Cube.AI automatically generates a new STM32-specialized model from raw data, replacing the need to train any model beforehand.

C Because STM32Cube.AI only supports tiny, single-sensor signal analysis models and cannot handle vision or multimodal models.

D Because STM32Cube.AI forces a full integer-only quantization of any model to guarantee smallest possible memory footprint.

E Because STM32Cube.AI only produces binary firmware blobs that can’t be inspected or integrated into custom projects.

What is the primary role of STM32Cube.AI's graph optimizer when preparing a neural network for an MCU?

A It rewrites and fuses operations and optimizes layout to improve runtime and resource use without changing the model’s functional logic.

B It prunes large portions of the network and removes layers until the model fits in RAM, even if accuracy drops significantly.

C It retrains the model on-device to better fit STM32 hardware characteristics before deployment.

D It converts all operations into platform-specific machine code instead of generating portable C code.

E It encrypts the model graph to prevent inspection and protect IP, which reduces memory usage.

Which statement best captures STM32Cube.AI's approach to quantization and how the tool helps you validate it?

A Cube.AI supports importing quantized models and can optionally quantize your model for you, but quantization modifies model logic so you should validate behavior—using your own data or a generated random stream if none is provided.

B Cube.AI only accepts float32 models and rejects any quantized formats to avoid changes in inference results.

C Cube.AI will always quantize every model to int8 losslessly without changing model behavior, so no validation is needed.

D Cube.AI performs quantization by training a special quantized network from scratch on ST's cloud, replacing your original model entirely.

E Cube.AI only supports binary (1-bit) quantization and refuses mixed or int8 quantization schemes.

How does the STM32Cube.AI developer cloud board farm help teams choose an appropriate MCU for deployment?

A By remotely flashing selected STM32 boards with the model, measuring inference time and per-layer memory use, and returning detailed benchmarks so teams can compare boards and trade-offs.

B By automatically selecting a production-grade MCU and locking the project to that single board with no benchmarking feedback.

C By providing only simulated timing estimates in the browser without running the model on real hardware, to avoid any need for physical boards.

D By training the model on ST's hardware farm to improve its accuracy before deployment to MCUs.

E By converting the model into a cloud-only service hosted by ST and preventing on-device execution.

Formatting help

italics	surround text with asterisks
bold	surround text with two asterisks
hyperlink	[hyperlink](https://example.com) or just a bare URL
code	surround text with `backticks`
~~strikethrough~~	surround text with ~~two tilde characters~~
quote	prefix with >

Upvotes Newest Oldest

mohammed.eshaq

Score: 1 | 3 years ago | no reply

We are unable to download the slides even.

mohammed.eshaq

Score: 1 | 3 years ago | no reply

Excellent tools and presentation! How can we download this session video for our future refference. It is very useful and important for the work that I do.

Partners & Sponsors

STMicroelectronics

Amazon AWS

Golioth

The Qt Company

Memfault

NXP

SEGGER Microcontroller

PE Micro

PX5

Become a Sponsor