Abstract

This publication describes a flexible, modular hardware system for running artificial intelligence (AI) and machine learning on battery-powered devices like smartwatches, fitness trackers, smart TVs, and Internet of Things (IoT) devices. Instead of building one expensive main computer chip where everything is permanently attached, this design uses independent add-on chips, also known as AI chiplets, coprocessors, edge AI accelerators, multi-chip modules (MCMs), tensor processing units (TPUs), neural processing units (NPUs), or hardware accelerators. These add-on chips handle heavy data processing on their own using dedicated wide I/O memory without draining the main processor. A software manager automatically detects how many of these add-on chips are attached via a high-speed die-to-die interface (or neural bus) and decides the most battery-efficient way to handle tasks. For example, if a user's smartwatch needs to constantly listen for a voice command or monitor their heart rate, the system can route these tasks to the add-on chips or to the cloud depending on the battery level. This modular approach saves power, keeps device costs down, and allows the exact same basic processor design to be used across both budget and premium electronic devices.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS