Speaker Seth Copen Goldstein
Affiliation Computer Science Department, Carnegie Mellon University
Host Stewart Tansley
Date recorded 30 July 2008
In this talk I will discuss a potential path towards building a massively connected machine with the processing power and connectivity similar to that of the human brain in one cubic meter. The construction method uses stochastic aggregation of modules to form a coherent highly-interconnected computer. We hope that the resulting system will provide a platform for performing brain related simulation and understanding algorithms and programming methods for massively connected computers.
The most interesting characteristic of our proposed machine is the interconnect. Preliminary simulations indicate that in the full scale machine each node will be able to communicate with over 100K other nodes. However, links are shared and the network behaves in a similar fashion to wireless networks with the associated advantages (broadcast is free) and disadvantages (hidden nodes and the need to negotiate for the use of a link). We have developed a working link protocol and have preliminary results on network routability. Using our communication primitives and the massive interconnectedness we have also developed a highly parallel version of the PageRank algorithm which depends on the connectivity of the machine.
©2008 Microsoft Corporation. All rights reserved.