Journal
IEEE OPEN JOURNAL OF CIRCUITS AND SYSTEMS
Volume 2, Issue -, Pages 782-791Publisher
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/OJCAS.2021.3116584
Keywords
Training; Circuits and systems; Memory architecture; Neural networks; Central Processing Unit; Sparse matrices; Hardware acceleration; Boltzmann machine; sparse matrix; FPGA; integer factorization
Categories
Funding
- JST Precursory Research for Embryonic Science and Technology (PRESTO) [JPMJPR18M5]
- Canon Medical Systems Corporation
Ask authors/readers for more resources
The paper introduces a scalable hardware architecture based on sparse Hamiltonian matrices and proposes a sparse matrix representation of PTELL for efficiently storing and computing Hamiltonians in hardware, achieving high-throughput operations with reduced memory usage.
Invertible logic has been recently presented that can realize bidirectional computing based on Hamiltonians for solving several critical issues, such as integer factorization and training neural networks. However, a hardware architecture for supporting large-scale general-purpose invertible logic has not been studied. In this paper, we introduce a scalable hardware architecture based on sparse Hamiltonian matrices. In order to store and compute the Hamiltonians efficiently in hardware, a sparse matrix representation of PTELL (partitioned and transposed ELLPACK) is proposed. A memory size of PTELL can be smaller than that of a conventional ELL by reducing the number of paddings while parallel reading of non-zero values are realized for high-throughput operations. As a result, PTELL achieves around 1% and 10% memory usages of a conventional dense and ELL matrices, respectively, in case of invertible multipliers. In addition, the proposed hardware accelerator of invertible logic for supporting arbitrary Hamiltonians is implemented on Xilinx VU9P FPGA, which achieves around two orders of magnitude faster than a 16-core Intel Xeon implementation.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available