Energy-Efficient Basecalling for ONT Long Reads on a Hybrid ASIC-GPU Platform

Cheng-You Tsai, Hung-Yu Tseng, Po-Yen Chang, Yi-Chang Lu

Abstract: This work presents an energy-efficient ASIC-GPU hybrid flow for basecalling sequences from raw current data generated using the third-generation sequencing machines designed by Oxford Nanopore Technologies. Compared to the official Bonito basecaller, the newly modified neural network has similar accuracy on nine representative testing bacterial genomes with only 28.5% of the model parameters. As a result, we can move our convolutional and recurrent layers to ASIC to save energy consumption. The proposed hybrid flow consists of an ASIC and an NVIDIA 2080 Ti GPU. Our design conducts convolution and matrix-vector multiplication operations efficiently. The ASICGPU flow can achieve the calling speed of $\mathbf{3 8 5 K}$ samples/sec with $2.57 \times$ and $2.05 \times$ improvement in throughput and energy efficiency compared to the pure GPU flow on an NVIDIA 3090.

External IDs:dblp:conf/isvlsi/TsaiTCL25