Configurable approximate multiplier for quantizing convolutional neural network and implementation method of configurable approximate multiplier

Abstract

The invention discloses a configurable approximate multiplier for quantizing a convolutional neural network and an implementation method of the configurable approximate multiplier. The configurable approximate multiplier comprises a symbol extension module, a sub multiplier module and an approximate adder. The symbol extension module splits long-bit-width signed fixed-point number multiplication into two short-bit-width signed fixed-point number multiplication. The sub-multiplier module comprises a plurality of sub-multipliers, each sub-multiplier only receives one signed fixed-point number output by the symbol extension module, and one signed fixed-point number multiplication is completed in combination with the other input; and the approximate adder merges results output by the sub-multiplier modules to obtain a final result of long-bit-width signed fixed-point number multiplication. For two signed fixed-point number multiplication operations with unequal input bit lengths, the speedand the energy efficiency are obviously improved; in a quantitative convolutional neural network with a large number of multiplication operations, the advantages of the method are embodied to the greatest extent.

Description

technical field [0001] The invention relates to engineering technical fields such as low power consumption design, approximate calculation, and convolutional neural network, and in particular to a configurable approximate multiplier for quantizing convolutional neural network and an implementation method thereof. Background technique [0002] Deep learning has achieved great success in the past few years due to its accuracy, robustness, and efficiency in various tasks. Deep learning typically employs convolutional neural network (CNN) architectures that can perform millions to billions of multiply and accumulate (MAC) operations per second. Deep learning is much more computationally intensive than traditional machine learning techniques. Therefore, energy efficiency (i.e., energy consumption per operation) has become key to deep learning implementation and deployment, especially for mobile and embedded devices that want to save energy and meet strict power constraints. [...

AI Technical Summary

Problems solved by technology

However, due to their diagonally shaped critical paths, most multiplier hardware circuits (cf. figure 1 in a) and figure 1 almost 75% in b)) are actually ineffective, thus wasting resources and compromising energy efficiency

Such DAS multipliers are even less efficient for quantized CNNs whose weights typically have a smaller bit width than the input, e.g. only 8 bits for 16 or 32 bit inputs

This asymmetry in operands with unequal bit widths makes array-based DAS multipliers not an ideal application

Images