Eugen Leitl wrote on Sat, 17 Dec 2011 10:43:09 +0100 > [300 EUR GPU] > [InfiniBand features]
Thanks for the tip about InfiniBand. I kept track of it while it was being developed but had wrongly assumed it had mostly died off when PCI Express started to become popular. It is actually a lot faster than my design in terms of bandwidth, though I think my latency is better (it has hard to compare since the architectures have some important differences). What I am doing it patching the Squeak VM so that the send bytecodes work the normal way when the receiver is in the same core as the sender but they become a "send message over the network" instruction if the receiver is remote. The Wikipedia article mentions an article about using a different topology (flattened butterfly) for a network using Infiniband in order to reduce the power consumed by the network. I was able to find the paper - "Energy proportional datacenter networks" by rs: Dennis Abts, Michael R. Marty, Philip M. Wells, Peter Klausler and Hong Liu. http://dl.acm.org/citation.cfm?id=1816004 > http://static.googleusercontent.com/external_content/untrusted_dlcp/research.google.com/pt-BR//pubs/archive/36462.pdf While looking for this, I saw that this conference's keynote was Chuck Thacker's Turing Award talk: > http://isca2010.inria.fr/media/slides/Turing-Improving_the_future_by_examining_the_past.pdf And this brings us right back to the start of this thread since he is saying the same thing that Alan Kay said at SJSU. -- Jecel _______________________________________________ fonc mailing list [email protected] http://vpri.org/mailman/listinfo/fonc
