Ruprecht-Karls-Universität Heidelberg


Design and Implementation of an EXTOLL Network-Interface for the Communication FPGA in the BrainScaleS Neuromorphic Computing System

Master Thesis by Tobias Thommes

Abstract:

Modern computer chips are constantly becoming more and more efficient, providing more Floating Point Operations per second (FLOPs) by consuming less or equal power than their predecessors. In spite of this trend, the power consumption of large supercomputers is still enormous: The number one on the Top 500 List, Sunway TaihuLight provides up to 125 PFLOPs and consumes power of around 15MW. The most efficient Super Computing System on the Green 500 List is Shoubu System B which provides 842 TFLOPs while consuming 50 kW. In contrast to that, the human brain can cope with intelligent operations and thoughts and additionally controls the human body, by only using an amount of about 20 W. The Human Brain Project (HBP) aims to understand by means of Synthesis Biology how this inconceivably efficient system works. The BrainScaleS system at the Kirchhof-Institute for Physics (KIP) in Heidelberg is part of the HBP and pursues this goal by developing a neuromorphic analog hardware system in combination with a conventional computing cluster. Up to now the BrainScaleS system is connected via Ethernet over USB 3.0 cables, which not only negatively affects latency and bandwidth, but also results in inefficient cabling density and effort. This work at hand describes the development of a new network interface for the FPGAs controlling the data communication between the neuromorphic hardware chips and the conventional digital system. The new interface will enable the BrainScaleS system to use the benefits of the EXTOLL network, a high-performance interconnection network, optimised for low latency and high message rates. This will significantly increase the network performance of the BrainScaleS system in terms of latency, message-rate and cabling effort.

 

« back

back to top