Research · Pioneer Research Institute 2025

Advancements in Hardware Accelerators for Sequence Alignment Algorithms

Under the mentorship of Hamilton College Professor Mark Bailey, I investigated how specialized hardware can accelerate sequence alignment, a computational foundation of genomic analysis.

I reviewed 13 accelerator designs, reimplemented two representative architectures in PyRTL, and wrote a 25-page research paper.

Finding speed at the software boundary

Next-generation sequencing produces vast numbers of short DNA fragments. Sequence alignment maps those reads to reference genomes, supporting applications from gene identification to phylogenetic analysis.

The volume of genomic data creates performance bottlenecks for conventional software. My research examines how FPGAs, ASICs, and processing-in-memory architectures address those limits.

13: hardware accelerator designs reviewed across recent literature
35: academic sources synthesized across architectures and biological applications.
2: representative architectures independently implemented and benchmarked in PyRTL
A: professor-awarded grade, #1 in cohort

I reimplemented an FPGA systolic-array design by El-Wafa et al. and a shift-register design by Di Tucci et al. Both ran on the same 4,096-base reference and 50 synthetic 512-base queries generated with a fixed random seed.

The shift-register architecture reached 511.75 cell updates per cycle, compared with 0.5 for the sequential baseline.

Logarithmic dot plot showing throughput of 0.5, 1, and 1.5 cell updates per cycle for El-Wafa variants, compared with 511.75 for the Di Tucci shift-register architecture.

The paper traces the field from FPGA systolic arrays through ASICs and processing-in-memory designs.

Across these designs, I identified seven recurring optimization methods.

Diagram aligning two DNA sequences and marking matches, a mismatch, and a gap.

Diagram showing sequence-alignment acceleration progressing from CPUs to FPGAs, ASICs, and processing-in-memory architectures.

Seven recurring accelerator optimizations: bit-level encoding, anti-diagonal parallelism, dynamic banding, windowing and tiling, lookahead and pipelining, traceback compression, and early termination.

Three future directions: biologically meaningful scoring, systems supporting both short and long reads, and standardized datasets and metrics.

A roadmap for faster genomic analysis

The review found a mapping between workflows and optimal designs. Banded Smith-Waterman and Bitap-based accelerators are suitable for short reads, while tile-based and wavefront approaches are more scalable for long reads.

I identify two hardware research directions: extending Bitap accelerators to support biologically meaningful scoring and designing general-purpose systems for both short and long reads. Standardized datasets and metrics are also needed for fair comparison.

Professor Mark Bailey, Chair of Computer Science at Hamilton College, awarded the paper an A and rated all four evaluated research qualities Excellent. He placed my potential for undergraduate academic work in the top 10% of students he has mentored and commented that the paper was "exceptional work."