Hybrid Dynamic Fixed Point Quantization Methodology for AI Accelerators.

Data quantization is an important task for the design of AI accelerators, because it has a great impact on both circuit area and inference accuracy. Dynamic fixed point (DFX) representation can save circuit area with accuracy loss, while dynamic dual-fixed-point (DDFX) representation can improve inference accuracy with area overhead. In this paper, we propose a hybrid dynamic fixed point (HDFX) quantization methodology: since the range of weight values is often limited, we use DFX to represent weight values; since the range of activation values is often large, we use DDFX to represent activation values. Based on the proposed HDFX quantization methodology, we develop the corresponding circuit for convolution operation. Experiment results show that the proposed methodology achieves good results in terms of both circuit area and inference accuracy.