Charge-domain CiM for Bayesian NNs

Uncertainty-aware quantized inference on charge-domain SRAM-CIM

A charge-domain SRAM compute-in-memory design targeting uncertainty-aware neural networks.

  • A variable-bit-precision neural network quantization method based on variational inference, targeting FLOP reduction.
  • A charge-domain SRAM-CIM for uncertainty-aware CNNs quantized by the above algorithm.
  • Energy reduction from charge reuse and unstructured sparsity routing / zero-skipping based on workload characteristics.