Helix: Algorithm/Architecture Co-design for Accelerating Nanopore Genome Base-calling

摘要: Nanopore genome sequencing is the key to enabling personalized medicine, global food security, and virus surveillance. The state-of-the-art base-callers adopt deep neural networks (DNNs) translate electrical signals generated by nanopore sequencers digital DNA symbols. A DNN-based base-caller consumes 44.5% of total execution time a pipeline. However, it difficult quantize build power-efficient processing-in-memory (PIM) run quantized base-caller. Although conventional network quantization techniques reduce computing overhead replacing floating-point multiply-accumulations cheaper fixed-point operations, significantly increases number systematic errors that cannot be corrected read votes. power density prior nonvolatile memory (NVM)-based PIMs has already exceeded thermal tolerance even with active heat sinks, because their efficiency severely limited analog-to-digital converters (ADC). Finally, Connectionist Temporal Classification (CTC) decoding voting cost 53.7% in base-caller, thus became its new bottleneck. In this paper, we propose novel algorithm/architecture co-designed PIM, Helix, power-efficiently accurately accelerate base-calling. From algorithm perspective, present error aware training minimize architecture low-power SOT-MRAM-based ADC array process conversion operations improve DNN PIMs. Moreover, revised traditional NVM-based dot-product engine CTC create SOT-MRAM binary comparator voting. Compared PIMs, Helix improves base-calling throughput 6x, per Watt 11.9x mm2 7.5x without degrading accuracy.