A prototype of 'Memister Chip,' a processing unit for machine learning that can operate at speeds up to 10,000 times faster than the CPU


Robert Coelius

The University of Michigan has announced that it has developed a new computing device that can perform machine learning at speeds up to 10,000 times faster than current CPUs installed on computers using resistors called 'Memorista'.

A fully integrated reprogrammable memristor – CMOS system for efficient multiply – accumulate operations | Nature Electronics

Firing up the synapses, the Memristor is now a proper computer – Blocks and Files

First programmable memristor computer aims to bring AI processing down from the cloud | University of Michigan News

A “ memister ” is a passive element that has the property of storing the passed charge and changing its resistance. The memristor has two roles of an arithmetic logic unit and a storage element. Since the storage element holds data in an analog form, it can be used as a non-volatile memory , and further, product-sum operations are possible.

The research team leader Wei Lu's 2013 announcement about Memorista can be read in the following articles:

A computer that can process images 1000 times faster than the conventional DAPRA with a ¥ 570 million funding support-GIGAZINE

The advantage of adopting memristor as a processing unit is that memristor can perform calculations suitable for neural networks, because it has the dual roles of memory and computing unit. 'GPU is about 10 times to about 100 times better than CPU in terms of power consumption and throughput,' said Lu, but Memesta chips may be about 10 times to about 100 times better than GPUs. Insists.

The research team actually manufactured a prototype model of the Memristor chip with 5800 Memristors,

OpenRISC CPU, communication circuit, and converter between analog and digital, and implemented a program to execute machine learning algorithm using Memristor chip. created. The following are the actually created memsta chips.

When the research team actually performed machine learning using this memory chip, it is 100% with basic perceptron that discriminates Greek letters and sparse modeling that identifies patterns of images and optimizes classification methods. Achieve accuracy. The two-layer neural network machine learning that finds common points and differentiating factors from breast cancer screening data has been able to classify malignant cancer and benign cancer with an accuracy of 94.6%.

However, Mr. Lu said that the analog information held inside Memister has a problem with reliability, and there are still issues to be used for commercial use, and will continue research to solve this problem. It says that there is.

in Hardware, Posted by darkhorse_log