A Hierarchical Multiple SIMD Architecture for Artificial Neural Networks

K.-S. Kim, C.-G. Kim, and S.-D. Kim (Korea)


Artificial Neural Network, Backpropagation learning, parallel algorithm, and Multiple SIMD.


This paper proposes a hierarchical multiple SIMD architecture (HMSA) for artificial neural networks. The HMSA consists of a global control unit, n local control units, and n2 processing elements interconnected with two-dimensional torus network. It provides an efficient PE control mechanism, which enables PEs to operate both on the row control mode (RC-mode) and column control mode (CC-mode) by supplying a particular operand vector concurrently to a group of PEs through the hierarchically configured control units (CU), i.e., a global CU and n local CUs. The local CUs are located at the diagonal position in torus configuration. ANNs algorithm which requires massive data parallel computation can be effectively mapped on the HMSA system. Performance evaluation is provided by detailed comparison in terms of the number of computation steps while mapping the multilayer perceptron with backpropagation learning to HMSA. The HMSA system can reduce 84.39%~89.56% of computation steps required by other architectures with corresponding algorithms.

Important Links:

Go Back