First Hybrid Memristor Chip Can Enable AI Processing Directly On Small Devices

  • Researchers for the first time develop fully functional, programmable memristor chip. 
  • It can enable AI processing directly on small, energy-constrained devices, including sensors and smartphones. 

Memristor is a compact device capable of both storing and processing information at the same physical location. These devices have been extensively studied for in-memory and neuromorphic applications.

Although engineers have already developed memristors for machine learning algorithms, they require external processing components to function. To achieve the optimal performance, the memristor crossbars must be well-integrated with peripheral and control circuitry.

Now, researchers at the University of Michigan for the first time have developed a fully functional programmable memristor chip that can enable AI processing directly on small, energy-constrained devices, including sensors and smartphones.

This means complex tasks such as voice commands would no longer have to be transferred to the cloud for interpretation. All computations could take place on your smartphone (without draining the battery quickly), speeding up response time and enhancing privacy and security.

Memristors Can Efficiently Implement AI Algorithms

Since memristors process and store data in the same location, they remove the barrier between processor and memory, significantly increasing computing speed. This is extremely helpful in implementing deep neural algorithms which deal with lots of data to perform complex tasks such as recognize objects in videos and images and identify the risk factor of serious diseases.

Developers prefer to run such algorithms on GPUs because they deliver better prediction accuracy and faster results at much lower costs compared to CPUs.

“GPUs are optimized for parallel processing and are 10-100 times better than CPUs in terms of power and throughput. The new memristor chip could be another 10-100 times better.” – Wei Lu, senior author of the paper.

Each memristor can perform thousands of operations [within a core] at once. The prototype developed in this study includes over 5,800 memristors. However, a commercial device would have million of memristors.

Reference: Nature Electronics | DOI:10.1038/s41928-019-0270-x | Michigan Engineering

Memristors arrays are specially designed to run machine learning algorithms that transform data into vectors (list of numerical representations). Usually, these vectors are stored in matrices which are mapped directly on the memristor arrays.

The First Programmable Memristor Chip

Programmable Memristor enable AI processingMemristor array chip plugs into the traditional computer chip | Credit: Robert Coelius, University of Michigan 

The research team first designed a chip to integrate the memristor array with other components — such as digital/analog converters, traditional digital process, and communication channels — required to program and run it.

They then integrated the memristor array on this chip and created a program to map machine learning algorithms onto the memristor array’s matrix-like structure.

They demonstrated the device with three different machine learning algorithms – a sparse coding algorithm, perceptron network and principal component analysis designed to identify patterns in complex data.

Read: 5 Quantum Processors That Feature New Computing Paradigm

The complete system can provide efficient solutions for various network sizes and applications in which real-time data processing and low energy consumption are critical factors.

Continued equipment, circuit, architecture innovations, and algorithm advances like quantized neural networks, can allow this memristor to handle far more complex and demanding tasks.

Written by
Varun Kumar

I am a professional technology and business research analyst with more than a decade of experience in the field. My main areas of expertise include software technologies, business strategies, competitive analysis, and staying up-to-date with market trends.

I hold a Master's degree in computer science from GGSIPU University. If you'd like to learn more about my latest projects and insights, please don't hesitate to reach out to me via email at [email protected].

View all articles
Leave a reply