A Mixed-Precision CNN–LSTM Accelerator Featuring Long-Kernel Memory Reuse and Dynamic Quantization for Sleep Staging

Abstract

This letter presents a mixed-precision CNN–LSTM accelerator for EEG-based sleep staging. To efficiently execute long temporal convolutions, a serpentine dataflow propagates activations between neighboring processing elements (PEs), reducing memory accesses by <inline-formula> <tex-math notation="LaTeX">$6.2\times $ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$2.25\times $ </tex-math></inline-formula> compared with output- and weight-stationary dataflows, respectively. A hybrid quantization scheme applies runtime dynamic quantization to the LSTM and multilayer perceptron (MLP) layers and assigns mixed precision on a layer-by-layer basis, reducing model size by <inline-formula> <tex-math notation="LaTeX">$4.16\times $ </tex-math></inline-formula>. Combined with a bit-flexible MAC architecture and channel-grouped memory scheduling, the accelerator supports INT2/4/8/16 inference. Fabricated in 65-nm CMOS, the accelerator consumes 0.40 mW and achieves a <inline-formula> <tex-math notation="LaTeX">$258\times $ </tex-math></inline-formula> speedup over an iso-frequency 32-b RISC-V processor. Evaluated across all 62 subjects in the public MASS-SS3 dataset using 31-fold leave-two-subject-out cross-validation, the mixed-precision approach with hybrid quantization maintains five-class accuracy within 0.15% of the floating-point baseline when optimized for general sleep staging and N3 sensitivity within 0.51% when optimized for deep-sleep detection. The developed techniques are broadly applicable to EEG-based deep learning (DL) workloads.