TinyML Talks - Demoing the World’s Fastest Inference Engine for Arm Cortex-M

This talk explains how an inference engine achieves faster, more memory-efficient execution of 8-bit deep-learning models on Arm Cortex-M microcontrollers. It covers memory planning, optimized and model-specific code, and benchmarks across publicly available neural-network models.

Syllabus

Intro
tinyML Summit 2022 Miniature dreams can come true. March 28-30, 2022 Hyatt Regency San Francisco Airport
You might know us from: Person detecti Person Presence Detection
Or from: the world's fastest Cortex-M inferen
How did we get here?
The machine learning flow
The tasks of an inference engine
An inference engine example: TELM
A closer look at the results
More off-the-shelf models
A closer look at the MLPerf Tiny models
How to beat the competition?
Memory planning: a (rotated) game of T
Memory planning for an example model
A much better memory plan
Even better: lower granularity planning
Memory planning at Plumerai: summary
Optimized INT8 code for speed
Model-specific code generation
The world's fastest Cortex-M inference
What can Plumerai mean for you?
Public benchmarking service: try it yours
Arm: The Software and Hardware Foundation for tin
EDGE IMPULSE The leading edge ML pla
Enabling the next generation of Sensor and Hearable pro to process rich data with energy efficiency
maxim integrated Maxim Integrated: Enabling Edge Intelligence
BROAD AND SCALABLE EDGE COMPUTING PORTFOLIO
SYNTIANT