Exploring the AMD® Deep Learning Processor Unit for Accelerating Selective Sweep Detection

Detecting selective sweeps, the genomic signature of positive selection, is fundamental to understanding evolutionary processes, and has many practical applications, such as identifying drug-resistant mutations in pathogens and designing more effective drug treatments. Deep Learning (DL) methods, particularly Convolutional Neural Networks (CNNs), have advanced the detection of selective sweeps by overcoming inherent theoretical limitations of traditional analytical techniques. Yet, the increased computational requirements of CNN classifiers employed within selective sweep detection frameworks hinder their practical deployment for large-scale, whole-genome scans. In this work, we explore the performance potential of the AMD^® Deep Learning Processing Unit (DPU) for accelerating a state-of-the-art CNN classifier that has been designed to distinguish selective sweeps from neutral genomic regions. The DPU is a programmable engine that is specifically designed for efficient CNN inference and can be integrated into the programmable logic of modern FPGAs through a software-centric development platform, thereby not requiring extensive computer architecture or hardware design background. Deploying three DPU instances onto a ZCU102 FPGA development board delivers up to 31.1x, 10.8x, and 24.1x higher throughput (classifications per second) than a single CPU core, 6 CPU cores, and a virtual GPU (Tesla P4) in the cloud, respectively. This exploratory study showcases the potential of software-programmable hardware acceleration for boosting performance of DL-based selective sweep detection, paving the way for efficient and practical whole-genome analyses.