» Current | 2020 | 2019 | 2018 | Subscribe

Linley Newsletter

Tsinghua Pedals Hybrid AI Processor

September 24, 2019

Author: Mike Demler

They say you never forget how to ride a bike, likely because learning to simultaneously balance, pedal, and steer takes a lot of practice. Led by Beijing’s Tsinghua University Center for Brain Inspired Computing Research (CBICR), a research team has trained a neural-network processor to handle all of those tasks, enabling an unmanned electric bicycle to navigate an obstacle course in response to voice commands.

This chip, called Tianjic (meaning "heavenly machine" in Chinese), implements a novel hybrid architecture that can concurrently run convolutional neural networks (CNNs), multilayer perceptrons (MLPs), recurrent neural networks (RNNs), and other deep neural networks (DNNs), as well as biologically inspired continuous-attractor and spiking neural networks (CANNs and SNNs).

Tianjic is the first chip flexible enough to natively run DNNs and SNNs, but owing to the necessary compromises, it’s best suited to modest workloads that combine both network types. Dedicated architectures designed for one or the other will always deliver superior performance per area and per watt. Voice-command recognition is a good fit for Tianjic’s SNN capabilities, and the ability to run small CNNs is useful for low-performance vision-based systems. The design isn’t a panacea that can unify all neural networks on a single architecture, but its ability to integrate CNNs, MLPs, and SNNs is a good fit for some robotics applications.

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Purchase the full article


Linley Spring Processor Conference 2020
Coming April 7-8, 2020
Hyatt Regency, Santa Clara, CA
Register Now!
Linley Fall Processor Conference 2020
Coming October 28-29, 2020
Hyatt Regency, Santa Clara, CA
More Events »


Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products
Subscribe to our Newsletter »