Processing-in-Storage Architecture for Machine Learning and Bioinformatics
TimeThursday, November 15th8:30am - 5pm
DescriptionUser-generated and bioinformatics database volumes has been increasing exponentially for more than a decade. With the slowdown and approaching end of Moore's law, traditional technologies cannot satisfy the increasing demands for processing power. This work presents PRINS, a highly-parallel in-storage processing architecture. PRINS combines non-volatile memory with processing capabilities on every bitcell. An emerging technology, memristors, form the basis for the design.
Implementations of three data-intensive and massively parallel algorithms are demonstrated: (1) Smith-Waterman DNA local sequence alignment (bioinformatics), (3) K-means clustering (machine learning) and (3) data deduplication. Performance and energy efficiency of PRINS compared to other published solutions is presented for each algorithm. PRINS is shown to achieve orders-of-magnitude improvement in performance and power efficiency over existing solutions, from large-scale bioinformatics and machine-learning to single-GPU or FPGA implementations.