In-Memory Accelerator Architectures for Machine Learning and Bioinformatics
TimeWednesday, November 14th8:30am - 5pm
DescriptionMost contemporary accelerators are von Neumann machines. With the increasing sizes of gathered and then processed data, memory bandwidth is the main limiting of performance. One approach to mitigate the bandwidth constraint is to bring the processing units closer to the data. This approach is known as near-data processing (NDP). However, NDP architecture (e.g., Hybrid Memory Cube) are still inherently limited because they are based on replicating the von Neumann architecture in memory.
My research proposes two new processing-in-storage architectures, where each bitcell can both store information and perform computation. The main building block of the architectures are memristors, an emerging memory technolgy.
The first architecture I propose, PRinS, was applied to accelerate machine learning and large-scale DNA sequence alignment. Using Associative Processing, PRinS achieves massive parallelism. The second, RASSA, accelerates DNA long read mapping, with approximate Hamming distance computation and quantifying mismatch of a pattern to voltage level.