A study led by Professor Lim Se-Min of the SATA Lab in the Department of Software Engineering at Kookmin University’s (President Jeong Seung Ryul) College of Software Convergence, conducted in collaboration with researchers from the University of California, Irvine (UC Irvine), was presented and published at ISCA 2026, the world’s most prestigious academic conference in the field of computer architecture. The paper, titled “Lembas: Cost-Efficient Genome Alignment with External Memory and FPGA Acceleration,” was presented at the 53rd ACM/IEEE International Symposium on Computer Architecture (ISCA 2026), held in Raleigh, North Carolina, from June 27 to July 1, and was included in the official proceedings. This is Kookmin University’s first ISCA paper, and it addresses a technology that reduces both the memory and computational costs required for large-scale genome analysis.

To perform a precise analysis of a genome, vast amounts of DNA sequences must be compared to identify matching positions and differences. As the volume of data to be analyzed increases, so does the burden on the memory required to store it and the equipment needed for computation. This study focused on two distinct bottlenecks in this process: in the initial stage of identifying matching regions between sequences, memory capacity is the limiting factor, while in the stage of calculating the specific correspondence between nucleotide sequences, computational throughput constrains performance and cost.
​
To address this, the research team developed “Lembas,” a genome alignment system that combines a CPU, an FPGA (a semiconductor that can reconfigure its circuitry), and an NVMe SSD (a high-speed storage device). Departing from conventional methods that required large-scale search data structures to be kept in memory during the initial stage, the system was designed to align data stored on the SSD using the FPGA and process it sequentially. During the stage of calculating alignment results, the team increased speed by processing the result path in tile units consisting of multiple cells, rather than retracing it one cell at a time. They also reduced the required hardware resources by configuring each processing stage to share the same FPGA over time.
​
The research team built a prototype by combining two AMD/Xilinx Alveo U50 FPGAs and four NVMe SSDs in a desktop-class computer and evaluated it using genome data from Arabidopsis thaliana, humans, and onions. System memory usage was reduced to about one-seventh of that of Minimap2, an existing genome alignment software. In reference alignment experiments using large-scale data generated from the onion genome, the system demonstrated processing performance up to 27 times faster than Minimap2 running on 64 threads and up to 2.4 times faster than a Parabricks system based on an NVIDIA A100 GPU. This achieved both lower memory requirements and a reduction in the total processing time for the actual alignment task.
​
Professor Lim Se-Min, the principal investigator of this study, stated, “As the scale of genomic analysis increases, so does the burden on memory and computing resources.” He added, “This study demonstrated that by designing data processing methods and hardware architectures tailored to the bottlenecks at each stage, we can reduce the resources and costs required for large-scale genomic analysis.” He continued, “We hope this will serve as a foundational technology supporting precision medicine and life sciences research by enabling more researchers to analyze large-scale genomic data even within limited computing environments.”