Open AccessDeep Architecture of Neural Networks Using Stochastic Integral Computation in VLSI
Author(s) -
B. Sumathy,
Kanagaraj Venusamy,
David Neels Pon Kumar D,
P. Malathi
Publication year - 2021
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1964/7/072012
Subject(s) - stochastic computing , computer science , implementation , field programmable gate array , latency (audio) , energy consumption , artificial neural network , very large scale integration , computation , architecture , computer architecture , range (aeronautics) , computer engineering , deep neural networks , embedded system , parallel computing , simple (philosophy) , distributed computing , artificial intelligence , algorithm , engineering , software engineering , telecommunications , art , philosophy , epistemology , aerospace engineering , electrical engineering , visual arts
Recently considerable attention has been paid to the deep neural networks many implementations currently need greater processes that match system architecture. However, multiple components and complex interconnections are usually needed, resulting in a wide range of occupational and abundant energy consumption. Stochastic technology clearly shows impressive outcomes for reduced area-efficient hardware implementations, although current stochastic techniques need lengthy flows that activate deep delays. We recommend a numerical model of stochastic computing in this paper and add a few simple circuits. We now suggest an effective execution of an integral SC-based DNN. The system framework was deployed on a Virtex7 Xilinx FPGA based on the average area and latency decrease of 45 % and 62 %, especially compared to the best validation design in the writings. We also consider semi-integration attributed to the fault-tolerant nature of stochastic structures, resulting in a 33 % reduction in energy consumption regarding the design of discrete datatype without even any compromising on efficiency.