Accelerating Selective Sweep Detection using AMD Deep Learning Processing Units and Vitis AI

Recent advancements in artificial intelligence have made the need for faster computation through AI acceleration increasingly important. This work explores using AMD’s Deep Learning Processor Units and Vitis AI to accelerate an image classification problem to identify a reduction in genetic variation, also known as selective sweep detection. An existing CNN designed for selective sweep detection is investigated and modified for faster image processing. The deployment of the original model showed slow processing due to several layers being handled by the CPU. By modifying the network and removing problematic layers, inference times improved significantly with minimal accuracy loss. The results suggest that DPUs are effective when models avoid custom or unsupported layers. Vitis AI simplifies quantization and compilation, but can be challenging to debug and requires specific software versions, limiting flexibility. Running the software on a cluster without root access also posed difficulties.