[Beowulf] Intel MXC

Mark Hahn hahn at mcmaster.ca
Fri Aug 16 10:46:10 PDT 2013


> Intel will introduce MXC, which they are calling "next generation
> optical connector" with speeds up to 1.6 Terabits per second at Intel
> Developer Forum (Sept. 10 to 12).

coverage of this is, as usual, very cacophonous.  everyone seems to 
just repeat the "smaller than rj45 or sfp" and 1.6 Tb factoids,
with a bit of hand-waving about previous Intel ejaculations on this
general topic (lightpeak, silicon photonics, etc).

> https://intel.activeevents.com/sf13/connect/search.ww#loadSearch-searchPhrase=MXC&searchType=session&tc=0&sortBy=abbreviationSort&p=

no idea how 25 Gb is derived from 1.6 Tb...

here's another rearrangement of the same ingredients:
http://www.extremetech.com/computing/163982-intel-teases-1-6tbps-optical-interconnect-asks-us-to-forget-about-light-peak

> Protocols supported include Infiniband, Ethernet and PCI-Express.

but that sounds like marketing gibberish.  I really don't think they've
implemented full-function controllers for those protocols, just that since 
those protocols use bits for signalling, they could go over this media ;)

> http://www.pcworld.com/article/2046676/intel-proposes-new-standard-to-light-up-data-transfers.html#tk.rss_all

I found the Insight64 comments completely uninsightful - this is really
only interesting if it moves quickly into systems, with cheapish cables
and affordable switches.  (actually, I wouldn't mind at all if Intel 
wanted to push a distributed switch model where each node has several 
ports and contains enough smarts to do simple routing.  there's no reason
this approach is incompatible with trendy stuff like SDN.  if we take 
the conventional wisdom that large non-HPC clusters (web, bigdata) 
do TOR aggregation, then a nic-peer fabric would make a lot of sense...)

regards, mark hahn.



More information about the Beowulf mailing list