[content] => 
    [params] => Array
            [0] => /forum/index.php?threads/survey-of-reram-memristor-based-designs-for-processing-in-memory-and-neural-network.10379/

    [addOns] => Array
            [DL6/MLTP] => 13
            [Hampel/JobRunner] => 1030170
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000670
            [ThemeHouse/XPress] => 1010394
            [XF] => 2010770
            [XFI] => 1030070

    [wordpress] => /var/www/html

Survey of ReRAM (memristor) based Designs for Processing-in-memory and Neural Network


New member
As data movement operations and power-budget become key bottlenecks in processor design, interest in approaches such as processing-in-memory (PIM), machine learning, and especially neural network (NN)-based accelerators has grown significantly. Resistive RAM (ReRAM or memristor) can work as both: high-density/low-energy storage and in-memory computation/search engine and hence, it is a promising technology.

present a survey of techniques for designing ReRAM-based PIM and NN architectures. As for PIM, the paper reviews use of memristor for implementing arithmetical operations, bitwise/logical operations and search operations, both exact and approximate. As for NN, the paper reviews use of memristor for accelerating both training and inference, and both convolution and fully-connected layers of CNN. The survey reviews 80+ papers.

PDF is here (could not upload the file here itself because its size exceed the file-size upload limit)


Not clear what the market is. Low power inferencing, like IBM NorthStar? The models are constantly evolving and signal processors offer flexible low power implementation of small and medium neural nets. These can be embedded in an ASIC or SOC with compatible processing. How would analog crospoints compete, how would they have the flexible topology? Interesting technology but how does it fit in?

Fred Chen

It seems the Processor-in-Memory and NN applications can be treated separately. NNs depend on analog weights while processors can be digital.

The other comment is that generalizing ReRAMs is not accurate. In particular, there is a mention of resistance drift. Some systems, like GST-based phase change memory, are well-known for this behavior, but otherwise, resistance drift is not regularly encountered.
Last edited: