A time-domain compute-in-memory architecture based on pulse width
modulation for multi-bits convolutional neural networks
Compute-in-memory (CIM) architecture is an effective way to reduce the
energy efficiency of convolutional neural networks (DNNs). This letter
presents a time-domain compute-in-memory design based on pulse width
modulation. Unique-weight convolution method is proposed for multi-bits
convolutional operations. A time-charge domain quantizer is also
presented to quantify the computation pulses of multi-rows in parallel.
Fabricated in 28nm CMOS technology, this design achieves 56.9% and
67.3% accuracy, 59.84Tops/W and 67.45Tops/W energy efficiency for 1-8-b
AlexNet and VGG16, respectively.