Non-volatile Content Addressable Memory for Computing Acceleration

Associative memory, as a form of computing-with-memory, reduces energy of the processing elements by eliminating redundant computations [2], [8], [11]–[15], [17]–[20]. An associative memory can quickly recall responses of a function for a subset of input patterns to save energy by avoiding the actual function execution on the processing element. An associative memory is typically composed of a ternary content-addressable memory (TCAM) to store input patterns and an output memory to return the pre-stored output. The operation of a TCAM goes beyond retrieving logic “0” and “1” and it has capability to store and search wildcard [5]. This feature opens the application of the TCAMs for approximate computing domain, and a wide range of applications in query processing [2], [14], text processing [26], search engine [9], [28], image processing [1], [7], [16], pattern recognition and classification [10], [21], [23]. Associative memories can be implemented on both software and hardware. Software solutions are based on hashing where a frequent data can be stored and retrieved from hash function by keys [3], [22]. In hardware, the associative memory is being implemented by TCAMs. Conventional TCAM with two SRAM cells suffers from high power dissipation and low density [?], [6]. The cost perbit of CMOS-based TCAM is 8× more than SRAM [4]. These limit the application of TCAMs to network applications [24]. Low leakage power of non-volatile memory make them appropriate as a replacement for CMOS-based TCAMs [29].