Abstract
Traditional Von Neumann computing is falling apart in the era of exploding data volumes as the overhead of data transfer becomes forbidding. Instead, it is more energy-efficient to fuse compute capability with memory where the data reside. This is particularly critical to pattern matching, a key computational step in large-scale data analytics, which involves repetitive search over very large databases residing in memory. Emerging spintronic technologies show remarkable versatility for the tight integration of logic and memory. In this article, we introduce SpinPM, a novel high-density, reconfigurable spintronic in-memory pattern matching spin–orbit torque (SOT)—specifically spin Hall effect (SHE)—substrate, and demonstrate the performance benefit SpinPM can achieve over conventional and near-memory processing systems.
Highlights
C LASSICAL computing platforms are not optimized for efficient data transfer, which complicates large-scale data analytics in the presence of exponentially growing data volumes
Improving the throughput performance in terms of number of patterns processed per second in an energy-efficient manner is especially challenging, considering that a representative reference can be around 109 characters long, at least 2 bits are necessary to encode each character, and a typical pattern data set can have hundreds of millions patterns to match [11], where SpinPM can help due to reduced data transfer overhead and parallel comparison/similarity score computations
BENCHMARKS We evaluate SpinPM using four pattern matching applications [which include common computational kernels for pattern matching such as bit count (BC)], besides the running example of deoxyribonucleic acid (DNA) sequence prealignment throughout this article
Summary
C LASSICAL computing platforms are not optimized for efficient data transfer, which complicates large-scale data analytics in the presence of exponentially growing data volumes. The most prevalent form is string matching via repetitive search over very large reference databases residing in memory Compute substrates, such as SpinPM, that collocate logic and memory to prevent slow and energy-hungry data transfers at scale, have great potential. In this case, each step of computation attempts to map a short character string to (the most similar substring of) an orders of magnitude longer character string and repeats this process for a very large number of short strings, where the longer string is fixed and acts as a reference. For each SpinPM input cell participating in computation, RWL is set to 1 to connect its MTJ with LL. For each SpinPM output cell participating in computation, WWL is set to 1 to turn the switch TM on, which in turn connects the SHE channel to the LL. While NOR gate is universal, we can implement different types of logic gates following a similar methodology for mapping the corresponding truth tables to the SpinPM array
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
