Ruprecht-Karls-Universität Heidelberg


Interfacing a Neuronal Accelerator to a High Performance Computing System

Bachelor Thesis by Fynn Beuttenmueller

Abstract:

The work at hand describes the connection of configurable neuromorphic computing substrates applying the means of EXTOLL technology. Within the subproject Neuromorphic Computing - as part of the overall Human Brain Project - research is performed in a hardware system that emulates accelerated neuronal networks. The high grade in connectivity and activity of neuronal models, as well as the speed-up factor of 103 to 104 in comparison to biological real time, are highly demanding with regard to bandwidth and latency for such a system. At first it is checked within this thesis work whether the EXTOLL technology is suitable for integration to the neuronal accelerator. A general survey on the accelerated neuromorphic computing system is performed and the system requirements towards an external data network are determined. Furthermore the EXTOLL technology with ist most current network board Tourmalet and its interface is presented. Latency and bandwidth in relevant network topologies are calculated with regard to neuronal networks and thereby the suitability of EXTOLL asserted. An EXTOLL interface module was designed and implemented, which is able to communicate neuronal pulse event information via an EXTOLL network. This interface connection has potential to be expanded exchanging configuration and system data in addition to the pulse event information via EXTOLL.

 

« back

back to top