(re)Assessing PiM Effectiveness for Sequence Alignment

Processing-in-Memory (PiM) technology has emerged as a promising solution for high-performance applications that hit the “memory wall”. This paper (re)examines PiM’s effectiveness in sequence alignment, which is a pivotal bottleneck in genome analysis. This application context ideally matches PiM’s strengths as it offers ample parallelism and is memory-intensive. We use commercially available PiM hardware (UPMEM) for an exploration based on direct implementation and performance analysis. Our results show that a minimally optimized implementation of a widely used alignment technique, Wavefront Alignment Algorithm (WFA) [18], outperforms highly optimized code targeting the UPMEM architecture by \(1.91\times \) and \(2.55\times \) , on average, in terms of throughput and power usage, respectively. Cost- and power-normalized comparisons suggest that, while the avenue offered by PiM is potentially promising, the existing technology is not yet mature enough to replace CPU platforms for sequence alignment.