Low-Latency Error-Prone Optical Networks for Fast Approximate Computation on High-End Datacenters

Cutting-edge application becomes deep learning, big-data processing, the approximate computation for NP-hard problems rather than exact scientific computation on large parallel computers including high-end datacenters and supercomputers. Such emerging applications are typically subject to various ki...

Full description

Saved in:
Bibliographic Details
Published in2019 24th OptoElectronics and Communications Conference (OECC) and 2019 International Conference on Photonics in Switching and Computing (PSC) pp. 1 - 2
Main Author Koibuchi, Michihiro
Format Conference Proceeding
LanguageEnglish
Japanese
Published The Institute of Electronics, Information and Communication Engineers (IEICE) 01.07.2019
Subjects
Approximate computing
Bandwidth
Forward error correction
Forward Error Correction (FEC)
interconnection networks
Modulation
Multiprocessor interconnection
optical networks
parallel computers
Supercomputers
Switches
Online AccessGet full text

Cover

More Information
Summary:Cutting-edge application becomes deep learning, big-data processing, the approximate computation for NP-hard problems rather than exact scientific computation on large parallel computers including high-end datacenters and supercomputers. Such emerging applications are typically subject to various kinds of numerical errors, which do not lead to execution failure. The author presents error-prone interconnection networks optimized for these emerging applications. Our interconnection networks provide high bandwidth and low latency at the sacrifice of accuracy. Existing interconnection networks provide almost error-free message transfer by detecting and correcting bit errors, while we do not. Interestingly, some emerging applications work on our error-prone interconnection network with a significant speedup.
DOI:10.23919/PS.2019.8817752