Dataflow Optimizations in a Sub-uW Data-Driven TCN Accelerator for Continuous ECG Monitoring

Vehicle Surroundings Perception Using Micro‐electromechanical Systems Inertial Sensors

“…Although the hardware accelerator in ref. [18] generally fits proposed CNN model, an appropriate degree of quantization needs to be found for fixed‐point arithmetic, while keeping classification accuracy high. In a first step, the softmax layer is removed from the full‐precision baseline, since it does not change the prediction result and omits logic for the non‐linear function.…”

“…The CNN architecture is inspired by the neural network architecture presented in ref. [18]. For the considered use‐case, it required key adjustments in the pre‐processing and feature extraction components.…”

“…pooling with stride 3. The kernel size and pool size parameters were chosen to fit the pre‐existing CNN accelerator, [ 18 ] while the number of filters were tuned using a parameter sweep. The outputs of the last convolutional layer are flattened and passed through a fully‐connected layer and softmax layer for classification.…”

“…Here, the hardware architecture from ref. [18] is used for evaluation. Note, that the choice of this implementation is mainly reasoned by both the close relation of the CNN model (see Table 4) and the stream processing use-case.…”

“…The measurement results of the fabricated chip in ref. [18] are used to calibrate the CNN accelerator power consumption of the proposed use-case. In this case, the word length and the number of MAC operations per prediction are used for scaling dynamic power, while static power is assumed to remain constant.…”