Signal Processing Methods to Enhance the Energy Efficiency of In-Memory Computing Architectures

Charbel Sakr, Naresh R. Shanbhag

Abstract: This paper presents signal processing methods to enhance the energy vs. accuracy trade-off of in-memory computing (IMC) architectures. First, an optimal clipping criterion (OCC) for signal quantization is proposed in order to minimize the precision of column analog-to-digital converters (ADCs) at iso-accuracy. For a Gaussian distributed signal, the OCC is shown to reduce the column ADC precision requirements by 3 bits at a signal-to-quantization noise ratio (SQNR) of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\text{22.5}\,dB$</tex-math></inline-formula> over the commonly used full range (FR) quantizer. Next, the input-sliced weight-parallel (ISWP) IMC architecture is presented as a generalization of the popular bit-serial bit-parallel (BSBP) architecture. Quantization noise analysis of the ISWP indicates that its accuracy is comparable to BSBP while providing an order-of-magnitude reduction in energy consumption due to fewer array invocations and smaller ADC precision. Combining OCC and ISWP noise analysis, we map popular DNNs such as VGG-9 (CIFAR-10), ResNet-18 (CIFAR-10), and AlexNet (ImageNet) on a OCC-enabled ISWP architecture and show a reduction in energy consumption by an order-of-magnitude at iso-accuracy over the BSBP architecture that employs FR-based ADCs.