Owen Densmore wrote:
We should be thinking thousands of processors with at least mesh
memory, if not hyper-cube.
That's fine if it is on a single piece of silicon or on the same
board. Latency across sites will be on the order of milliseconds just
from the speed of light. Compare to nanoseconds for intra-chip
communication. Even state-of-the-art interconnects for installations
like Encanto are still just on the order of microseconds within the
facility.
Google will eventually beat all these efforts because they are
thinking plumbing/networking with scalable data stores (NoSql).
Hmm, I think their application toolkits will get good at use cases where
there are medium and high latencies to deal with along with medium
individual bandwidths, e.g. JavaScript within a web browser and delays
from communication over the internet. And they'll get better and better
at managing millions of such workloads. That's completely different
from high performance computing and scientific workloads.
Marcus
============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org
