Answers are grounded in this paper and cite their sources.
In this paper, we consider possibility of using matrix multipliers in hardware implementation of convolutional neural networks. Literature review has been conducted, existing matrix multiplication methods and their hardware implementations has been analyzed. Systolic algorithm for matrix multiplication (SMM) based on divide-and-conquer approach is proposed. This algorithm has been implemented as parameterized generator of hardware descriptions. Its positive and negative aspects are identified. Main difficulties of preparing feature maps for matrix multiplication are analyzed. Comparative analysis of hardware implementations of MobileNet neural network has been carried out for implementations based on convolutional blocks and implementations based on systolic matrix multiplier. A combined method is proposed that has the advantages of both approaches.
No takes yet. Share an insight, caveat, or question.
Stempkovskiy et al. (2023) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: