» Current | 2021 | 2020 | 2019 | Subscribe

Linley Newsletter

IBM Trains in Analog to Save AI Power

August 14, 2018

Author: Mike Demler

Neural networks consume most of their power moving data to and from memory. To ameliorate this problem, researchers at IBM and other companies are developing methods for performing some calculations directly in memory. IBM is using phase-change memory (PCM), a nonvolatile storage mechanism that enables high-precision analog multiplication. It employs storage cells composed of materials that change conductance by as much as four orders of magnitude when heated over a large temperature range.

At this stage of its PCM research, IBM’s primary goal is to demonstrate that the technology can match the accuracy of neural networks trained using popular digital methods, while consuming less power. Although the company’s comparison with a TensorFlow model was completely in software, its test results accomplished that goal. The PCM-based network matched TensorFlow’s accuracy on the MNIST database of handwritten characters and the CIFAR-10 image data set, and it was within 1% on CIFAR-100.

Nevertheless, performance is critical to neural-network training as well, since huge data sets can take hours to process. We expect the PCM technique is more likely to succeed in embedded applications. The nonvolatile memories are also ideal for fixed-function inference engines, since they retain their network calibration even after losing power. To take advantage of the area and power efficiencies of its design, IBM should focus on low-power IoT clients and similar applications, which could benefit from retraining in the field using newly acquired sensor data.

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

Free Newsletter

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


Linley Fall Processor Conference 2021
Coming October 20-21, 2021
Hyatt Regency Hotel, Santa Clara, CA
Register Now!
Linley Spring Processor Conference 2021
April 19 - 23, 2021
Proceedings Available
More Events »